Daeyoung Kim

Jungsoo Lee, Jeonghoon Park, Daeyoung Kim et al.

In image classification, "debiasing" aims to train a classifier to be less susceptible to dataset bias, the strong correlation between peripheral attributes of data samples and a target class. For example, even if the frog class in the dataset mainly consists of frog images with a swamp background (i.e., bias-aligned samples), a debiased classifier should be able to correctly classify a frog at a beach (i.e., bias-conflicting samples). Recent debiasing approaches commonly use two components for debiasing, a biased model $f_B$ and a debiased model $f_D$. $f_B$ is trained to focus on bias-aligned samples (i.e., overfitted to the bias) while $f_D$ is mainly trained with bias-conflicting samples by concentrating on samples which $f_B$ fails to learn, leading $f_D$ to be less susceptible to the dataset bias. While the state-of-the-art debiasing techniques have aimed to better train $f_D$, we focus on training $f_B$, an overlooked component until now. Our empirical analysis reveals that removing the bias-conflicting samples from the training set for $f_B$ is important for improving the debiasing performance of $f_D$. This is due to the fact that the bias-conflicting samples work as noisy samples for amplifying the bias for $f_B$ since those samples do not include the bias attribute. To this end, we propose a simple yet effective data sample selection method which removes the bias-conflicting samples to construct a bias-amplified dataset for training $f_B$. Our data sample selection method can be directly applied to existing reweighting-based debiasing approaches, obtaining consistent performance boost and achieving the state-of-the-art performance on both synthetic and real-world datasets.

Seongjun Yang, Hyeonji Hwang, Daeyoung Kim et al.

Federated learning (FL) is an active area of research. One of the most suitable areas for adopting FL is the medical domain, where patient privacy must be respected. Previous research, however, does not provide a practical guide to applying FL in the medical domain. We propose empirical benchmarks and experimental settings for three representative medical datasets with different modalities: longitudinal electronic health records, skin cancer images, and electrocardiogram signals. The likely users of FL such as medical institutions and IT companies can take these benchmarks as guides for adopting FL and minimize their trial and error. For each dataset, each client data is from a different source to preserve real-world heterogeneity. We evaluate six FL algorithms designed for addressing data heterogeneity among clients, and a hybrid algorithm combining the strengths of two representative FL algorithms. Based on experiment results from three modalities, we discover that simple FL algorithms tend to outperform more sophisticated ones, while the hybrid algorithm consistently shows good, if not the best performance. We also find that a frequent global model update leads to better performance under a fixed training iteration budget. As the number of participating clients increases, higher cost is incurred due to increased IT administrators and GPUs, but the performance consistently increases. We expect future users will refer to these empirical benchmarks to design the FL experiments in the medical domain considering their clinical tasks and obtain stronger performance with lower costs.

Ji-Hoon Hwang, Daeyoung Kim, Hyung-Suk Yoon et al.

Semantic segmentation is crucial for autonomous navigation in off-road environments, enabling precise classification of surroundings to identify traversable regions. However, distinctive factors inherent to off-road conditions, such as source-target domain discrepancies and sensor corruption from rough terrain, can result in distribution shifts that alter the data differently from the trained conditions. This often leads to inaccurate semantic label predictions and subsequent failures in navigation tasks. To address this, we propose ST-Seg, a novel framework that expands the source distribution through style expansion (SE) and texture regularization (TR). Unlike prior methods that implicitly apply generalization within a fixed source distribution, ST-Seg offers an intuitive approach for distribution shift. Specifically, SE broadens domain coverage by generating diverse realistic styles, augmenting the limited style information of the source domain. TR stabilizes local texture representation affected by style-augmented learning through a deep texture manifold. Experiments across various distribution-shifted target domains demonstrate the effectiveness of ST-Seg, with substantial improvements over existing methods. These results highlight the robustness of ST-Seg, enhancing the real-world applicability of semantic segmentation for off-road navigation.

Daeyoung Kim, Seongsu Bae, Seungho Kim et al.

Question Answering on Electronic Health Records (EHR-QA) has a significant impact on the healthcare domain, and it is being actively studied. Previous research on structured EHR-QA focuses on converting natural language queries into query language such as SQL or SPARQL (NLQ2Query), so the problem scope is limited to pre-defined data types by the specific query language. In order to expand the EHR-QA task beyond this limitation to handle multi-modal medical data and solve complex inference in the future, more primitive systemic language is needed. In this paper, we design the program-based model (NLQ2Program) for EHR-QA as the first step towards the future direction. We tackle MIMICSPARQL*, the graph-based EHR-QA dataset, via a program-based approach in a semi-supervised manner in order to overcome the absence of gold programs. Without the gold program, our proposed model shows comparable performance to the previous state-of-the-art model, which is an NLQ2Query model (0.9% gain). In addition, for a reliable EHR-QA model, we apply the uncertainty decomposition method to measure the ambiguity in the input question. We empirically confirmed data uncertainty is most indicative of the ambiguity in the input question.

Keon Lee, Kyumin Park, Daeyoung Kim

The majority of current Text-to-Speech (TTS) datasets, which are collections of individual utterances, contain few conversational aspects. In this paper, we introduce DailyTalk, a high-quality conversational speech dataset designed for conversational TTS. We sampled, modified, and recorded 2,541 dialogues from the open-domain dialogue dataset DailyDialog inheriting its annotated attributes. On top of our dataset, we extend prior work as our baseline, where a non-autoregressive TTS is conditioned on historical information in a dialogue. From the baseline experiment with both general and our novel metrics, we show that DailyTalk can be used as a general TTS dataset, and more than that, our baseline can represent contextual information from DailyTalk. The DailyTalk dataset and baseline code are freely available for academic use with CC-BY-SA 4.0 license.

Junsoo Park, Seungyeon Jwa, Meiying Ren et al.

Employing Large Language Models (LLMs) to assess the quality of generated responses, such as prompting instruct-tuned models or fine-tuning judge models, has become a widely adopted evaluation method. It is also known that such evaluators are vulnerable to biases, such as favoring longer responses. While it is important to overcome this problem, the specifics of these biases remain under-explored. In this work, we qualitatively identify six types of biases inherent in various judge models. We propose EvalBiasBench as a meta-evaluation collection of hand-crafted test cases for each bias type. Additionally, we present de-biasing dataset construction methods and the associated preference dataset OffsetBias. Experimental results demonstrate that fine-tuning on our dataset significantly enhances the robustness of judge models against biases and improves performance across most evaluation scenarios. We release our datasets and the fine-tuned judge model to public.

Hankyu Jang, Daeyoung Kim

Neural radiance field (NeRF) attracts attention as a promising approach to reconstructing the 3D scene. As NeRF emerges, subsequent studies have been conducted to model dynamic scenes, which include motions or topological changes. However, most of them use an additional deformation network, slowing down the training and rendering speed. Tensorial radiance field (TensoRF) recently shows its potential for fast, high-quality reconstruction of static scenes with compact model size. In this paper, we present D-TensoRF, a tensorial radiance field for dynamic scenes, enabling novel view synthesis at a specific time. We consider the radiance field of a dynamic scene as a 5D tensor. The 5D tensor represents a 4D grid in which each axis corresponds to X, Y, Z, and time and has 1D multi-channel features per element. Similar to TensoRF, we decompose the grid either into rank-one vector components (CP decomposition) or low-rank matrix components (newly proposed MM decomposition). We also use smoothing regularization to reflect the relationship between features at different times (temporal dependency). We conduct extensive evaluations to analyze our models. We show that D-TensoRF with CP decomposition and MM decomposition both have short training times and significantly low memory footprints with quantitatively and qualitatively competitive rendering results in comparison to the state-of-the-art methods in 3D dynamic scene modeling.

Kyumin Park, Keon Lee, Daeyoung Kim et al.

Even with recent advances in speech synthesis models, the evaluation of such models is based purely on human judgement as a single naturalness score, such as the Mean Opinion Score (MOS). The score-based metric does not give any further information about which parts of speech are unnatural or why human judges believe they are unnatural. We present a novel speech dataset, RedPen, with human annotations on unnatural speech regions and their corresponding reasons. RedPen consists of 180 synthesized speeches with unnatural regions annotated by crowd workers; These regions are then reasoned and categorized by error types, such as voice trembling and background noise. We find that our dataset shows a better explanation for unnatural speech regions than the model-driven unnaturalness prediction. Our analysis also shows that each model includes different types of error types. Summing up, our dataset successfully shows the possibility that various error regions and types lie under the single naturalness score. We believe that our dataset will shed light on the evaluation and development of more interpretable speech models in the future. Our dataset will be publicly available upon acceptance.

Son Nguyen, Mikel Lainsa, Hung Dao et al.

Occlusion is a long-standing problem in computer vision, particularly in instance segmentation. ACM MMSports 2023 DeepSportRadar has introduced a dataset that focuses on segmenting human subjects within a basketball context and a specialized evaluation metric for occlusion scenarios. Given the modest size of the dataset and the highly deformable nature of the objects to be segmented, this challenge demands the application of robust data augmentation techniques and wisely-chosen deep learning architectures. Our work (ranked 1st in the competition) first proposes a novel data augmentation technique, capable of generating more training samples with wider distribution. Then, we adopt a new architecture - Hybrid Task Cascade (HTC) framework with CBNetV2 as backbone and MaskIoU head to improve segmentation performance. Furthermore, we employ a Stochastic Weight Averaging (SWA) training strategy to improve the model's generalization. As a result, we achieve a remarkable occlusion score (OM) of 0.533 on the challenge dataset, securing the top-1 position on the leaderboard. Source code is available at this https://github.com/nguyendinhson-kaist/MMSports23-Seg-AutoID.

Jimin Hong, Jungsoo Park, Daeyoung Kim et al.

With contrastive pre-training, sentence encoders are generally optimized to locate semantically similar samples closer to each other in their embedding spaces. In this work, we focus on the potential of their embedding spaces to be readily adapted to zero-shot text classification, as semantically distinct samples are already well-separated. Our framework, RaLP (Retrieval augmented Label Prompts for sentence encoder), encodes prompted label candidates with a sentence encoder, then assigns the label whose prompt embedding has the highest similarity with the input text embedding. In order to compensate for the potentially poorly descriptive labels in their original format, RaLP retrieves sentences that are semantically similar to the original label prompt from external corpora and use them as additional pseudo-label prompts. RaLP achieves competitive or stronger performance than much larger baselines on various closed-set classification and multiple-choice QA datasets under zero-shot settings. We show that the retrieval component plays a pivotal role in RaLP's success, and its results are robustly attained regardless of verbalizer variations.

Jeongho Ju, Daeyoung Kim, SunYoung Park et al.

In this paper, we introduce an open-source Korean-English vision-language model (VLM), VARCO-VISION. We incorporate a step-by-step training strategy that allows a model learn both linguistic and visual information while preserving the backbone model's knowledge. Our model demonstrates outstanding performance in diverse settings requiring bilingual image-text understanding and generation abilities compared to models of similar size. VARCO-VISION is also capable of grounding, referring, and OCR, expanding its usage and potential applications for real-world scenarios. In addition to the model, we release five Korean evaluation datasets, including four closed-set and one openset benchmarks. We anticipate that our milestone will broaden the opportunities for AI researchers aiming to train VLMs. VARCO-VISION is available at https://huggingface.co/NCSOFT/VARCO-VISION-14B.

Thao Do, Dinh Phu Tran, An Vo et al.

Extracting fine-grained OCR text from aged documents in diacritic languages remains challenging due to unexpected artifacts, time-induced degradation, and lack of datasets. While standalone spell correction approaches have been proposed, they show limited performance for historical documents due to numerous possible OCR error combinations and differences between modern and classical corpus distributions. We propose a method utilizing available content-focused ebooks as a reference base to correct imperfect OCR-generated text, supported by large language models. This technique generates high-precision pseudo-page-to-page labels for diacritic languages, where small strokes pose significant challenges in historical conditions. The pipeline eliminates various types of noise from aged documents and addresses issues such as missing characters, words, and disordered sequences. Our post-processing method, which generated a large OCR dataset of classical Vietnamese books, achieved a mean grading score of 8.72 on a 10-point scale. This outperformed the state-of-the-art transformer-based Vietnamese spell correction model, which scored 7.03 when evaluated on a sampled subset of the dataset. We also trained a baseline OCR model to assess and compare it with well-known engines. Experimental results demonstrate the strength of our baseline model compared to widely used open-source solutions. The resulting dataset will be released publicly to support future studies.

Dinh Phu Tran, Dao Duy Hung, Daeyoung Kim

Recently, window-based attention methods have shown great potential for computer vision tasks, particularly in Single Image Super-Resolution (SISR). However, it may fall short in capturing long-range dependencies and relationships between distant tokens. Additionally, we find that learning on spatial domain does not convey the frequency content of the image, which is a crucial aspect in SISR. To tackle these issues, we propose a new Channel-Partitioned Attention Transformer (CPAT) to better capture long-range dependencies by sequentially expanding windows along the height and width of feature maps. In addition, we propose a novel Spatial-Frequency Interaction Module (SFIM), which incorporates information from spatial and frequency domains to provide a more comprehensive information from feature maps. This includes information about the frequency content and enhances the receptive field across the entire image. Experimental findings show the effectiveness of our proposed modules and architecture. In particular, CPAT surpasses current state-of-the-art methods by up to 0.31dB at x2 SR on Urban100.

Saad Wazir, Daeyoung Kim

Segmenting biomarkers in medical images is crucial for various biotech applications. Despite advances, Transformer and CNN based methods often struggle with variations in staining and morphology, limiting feature extraction. In medical image segmentation, where datasets often have limited sample availability, recent state-of-the-art (SOTA) methods achieve higher accuracy by leveraging pre-trained encoders, whereas end-to-end methods tend to underperform. This is due to challenges in effectively transferring rich multiscale features from encoders to decoders, as well as limitations in decoder efficiency. To address these issues, we propose an architecture that captures multi-scale local and global contextual information and a novel decoder design, which effectively integrates features from the encoder, emphasizes important channels and regions, and reconstructs spatial dimensions to enhance segmentation accuracy. Our method, compatible with various encoders, outperforms SOTA methods, as demonstrated by experiments on four datasets and ablation studies. Specifically, our method achieves absolute performance gains of 2.76% on MoNuSeg, 3.12% on DSB, 2.87% on Electron Microscopy, and 4.03% on TNBC datasets compared to existing SOTA methods. Code: https://github.com/saadwazir/MCADS-Decoder

Saad Wazir, Daeyoung Kim

Identifying biomarkers in medical images is vital for a wide range of biotech applications. However, recent Transformer and CNN based methods often struggle with variations in morphology and staining, which limits their feature extraction capabilities. In medical image segmentation, where data samples are often limited, state-of-the-art (SOTA) methods improve accuracy by using pre-trained encoders, while end-to-end approaches typically fall short due to difficulties in transferring multiscale features effectively between encoders and decoders. To handle these challenges, we introduce a nested UNet architecture that captures both local and global context through Multiscale Feature Fusion and Attention Mechanisms. This design improves feature integration from encoders, highlights key channels and regions, and restores spatial details to enhance segmentation performance. Our method surpasses SOTA approaches, as evidenced by experiments across four datasets and detailed ablation studies. Code: https://github.com/saadwazir/ReN-UNet

Seongah Kim, Dinh Phu Tran, Hyeontaek Hwang et al.

Audio-visual understanding requires effective alignment between heterogeneous modalities, yet cross-modal correspondence remains challenging when temporally aligned audio and visual signals lack clear semantic correspondence.We propose to use text as a semantic anchor for audio-visual representation learning.To this end, we introduce a parameter-efficient adaptation frameworkbuilt on frozen audio and visual encoders, centered on Text-Bridged Audio-Visual Adapter (TB-AVA), which enables text-mediated interaction between audio and visual streams. At the core of TB-AVA, Gated Semantic Modulation (GSM) selectively modulates feature channels based on text-inferred semantic relevance. We evaluate the proposed approach on multiple benchmarks, including AVE, AVS, and AVVP, where the proposed framework achieves state-of-the-art performance, demonstrating text as an effective semantic anchor for parameter-efficient fine-tuning (PEFT) in audio-visual learning.

Dinh Phu Tran, Jihoon Jeong, Saad Wazir et al.

We present a formal problem formulation for \textit{Reliable} Audio-Visual Question Answering ($\mathcal{R}$-AVQA), where we prefer abstention over answering incorrectly. While recent AVQA models have high accuracy, their ability to identify when they are likely wrong and their consequent abstention from answering remain underexplored areas of research. To fill this gap, we explore several approaches and then propose Adaptive Confidence Refinement (ACR), a lightweight method to further enhance the performance of $\mathcal{R}$-AVQA. Our key insight is that the Maximum Softmax Probability (MSP) is Bayes-optimal only under strong calibration, a condition usually not met in deep neural networks, particularly in multimodal models. Instead of replacing MSP, our ACR maintains it as a primary confidence signal and applies input-adaptive residual corrections when MSP is deemed unreliable. ACR introduces two learned heads: i) a Residual Risk Head that predicts low-magnitude correctness residuals that MSP does not capture, and ii) a Confidence Gating Head to determine MSP trustworthiness. Our experiments and theoretical analysis show that ACR consistently outperforms existing methods on in- and out-of-disrtibution, and data bias settings across three different AVQA architectures, establishing a solid foundation for $\mathcal{R}$-AVQA task. The code and checkpoints will be available upon acceptance \href{https://github.com/PhuTran1005/R-AVQA}{at here}

Vy Tuong Dang, An Vo, Quang Tau et al.

Vision language models (VLMs) demonstrate remarkable capabilities on English multimodal tasks, but their performance on low-resource languages with genuinely multimodal educational content remains largely unexplored. In this work, we test how VLMs perform on Vietnamese educational assessments, investigating whether VLMs trained predominantly on English data can handle real-world cross-lingual multimodal reasoning. Our work presents the first comprehensive evaluation of VLM capabilities on multimodal Vietnamese exams through proposing ViExam, a benchmark containing 2,548 multimodal questions. We find that state-of-the-art VLMs achieve only 57.74% while open-source models achieve 27.70% mean accuracy across 7 academic domains, including Mathematics, Physics, Chemistry, Biology, Geography, Driving Test, and IQ Test. Most VLMs underperform average human test-takers (66.54%), with only the thinking VLM o3 (74.07%) exceeding human average performance, yet still falling substantially short of human best performance (99.60%). Cross-lingual prompting with English instructions while maintaining Vietnamese content fails to improve performance, decreasing accuracy by 1 percentage point for SOTA VLMs. Human-in-the-loop collaboration can partially improve VLM performance by 5 percentage points. Code and data are available at: https://vi-exam.github.io.

Keon Lee, Kyumin Park, Daeyoung Kim

Previous works on neural text-to-speech (TTS) have been addressed on limited speed in training and inference time, robustness for difficult synthesis conditions, expressiveness, and controllability. Although several approaches resolve some limitations, there has been no attempt to solve all weaknesses at once. In this paper, we propose STYLER, an expressive and controllable TTS framework with high-speed and robust synthesis. Our novel audio-text aligning method called Mel Calibrator and excluding autoregressive decoding enable rapid training and inference and robust synthesis on unseen data. Also, disentangled style factor modeling under supervision enlarges the controllability in synthesizing process leading to expressive TTS. On top of it, a novel noise modeling pipeline using domain adversarial training and Residual Decoding empowers noise-robust style transfer, decomposing the noise without any additional label. Various experiments demonstrate that STYLER is more effective in speed and robustness than expressive TTS with autoregressive decoding and more expressive and controllable than reading style non-autoregressive TTS. Synthesis samples and experiment results are provided via our demo page, and code is available publicly.

Trung Quang Tran, Mingu Kang, Daeyoung Kim

This paper proposes integrating semantics-oriented similarity representation into RankingMatch, a recently proposed semi-supervised learning method. Our method, dubbed ReRankMatch, aims to deal with the case in which labeled and unlabeled data share non-overlapping categories. ReRankMatch encourages the model to produce the similar image representations for the samples likely belonging to the same class. We evaluate our method on various datasets such as CIFAR-10, CIFAR-100, SVHN, STL-10, and Tiny ImageNet. We obtain promising results (4.21% error rate on CIFAR-10 with 4000 labels, 22.32% error rate on CIFAR-100 with 10000 labels, and 2.19% error rate on SVHN with 1000 labels) when the amount of labeled data is sufficient to learn semantics-oriented similarity representation. The code is made publicly available at https://github.com/tqtrunghnvn/ReRankMatch.

Trung Q. Tran, Giang V. Nguyen, Daeyoung Kim

Human pose estimation - the process of recognizing human keypoints in a given image - is one of the most important tasks in computer vision and has a wide range of applications including movement diagnostics, surveillance, or self-driving vehicle. The accuracy of human keypoint prediction is increasingly improved thanks to the burgeoning development of deep learning. Most existing methods solved human pose estimation by generating heatmaps in which the ith heatmap indicates the location confidence of the ith keypoint. In this paper, we introduce novel network structures referred to as multi-resolution representation learning for human keypoint prediction. At different resolutions in the learning process, our networks branch off and use extra layers to learn heatmap generation. We firstly consider the architectures for generating the multi-resolution heatmaps after obtaining the lowest-resolution feature maps. Our second approach allows learning during the process of feature extraction in which the heatmaps are generated at each resolution of the feature extractor. The first and second approaches are referred to as multi-resolution heatmap learning and multi-resolution feature map learning respectively. Our architectures are simple yet effective, achieving good performance. We conducted experiments on two common benchmarks for human pose estimation: MSCOCO and MPII dataset. The code is made publicly available at https://github.com/tqtrunghnvn/SimMRPose.

Thao Do, Dinh Phu Tran, An Vo et al.

Efficient context compression is crucial for improving the accuracy and scalability of question answering. For the efficiency of Retrieval Augmented Generation, context should be delivered fast, compact, and precise to ensure clue sufficiency and budget-friendly LLM reader cost. We propose a margin-based framework for query-driven context pruning, which identifies sentences that are critical for answering a query by measuring changes in clue richness when they are omitted. The model is trained with a composite ranking loss that enforces large margins for critical sentences while keeping non-critical ones near neutral. Built on a lightweight encoder-only Transformer, our approach generally achieves strong exact-match and F1 scores with high-throughput inference and lower memory requirements than those of major baselines. In addition to efficiency, our method yields effective compression ratios without degrading answering performance, demonstrating its potential as a lightweight and practical alternative for retrieval-augmented tasks.

Duc Dm, Thao Do, Minh Son Hoang et al.

Differentially private (DP) training protects individual examples by adding noise to gradients, but the injected noise interacts nontrivially with adaptive optimizers. Recent DP methods temporally filter privatized gradients to reduce variance; however, filtering also changes the DP noise statistics seen by AdamW's second-moment accumulator. As a result, bias corrections derived for unfiltered DP noise, such as subtracting sigma_w squared, can become miscalibrated when filtering is present. We propose FiBeR, a DP optimizer designed for temporally filtered privatized gradients. FiBeR (i) performs denoising in innovation space by filtering the residual stream and integrating it to form the filtered gradient estimate, (ii) decouples the two-point observation geometry from the innovation gain to enable independent tuning, and (iii) introduces a filter-aware second-moment calibration that subtracts the attenuated DP noise contribution A(omega) sigma_w squared, where A(omega) is derived in closed form for the innovation filter and can be computed for general stable linear filters. Across vision and language benchmarks, FiBeR consistently demonstrates substantial improvements in the performance of DP optimizers, surpassing state-of-the-art results under equivalent privacy constraints on multiple tasks.

An Vo, Khai-Nguyen Nguyen, Mohammad Reza Taesiri et al.

Large language models (LLMs) memorize a vast amount of prior knowledge from the Internet that helps them on downstream tasks but also may notoriously sway their outputs towards wrong or biased answers. In this work, we test how the knowledge about popular subjects hurt the accuracy of vision language models (VLMs) on standard, objective visual tasks of counting and identification. We find that state-of-the-art VLMs are strongly biased (e.g., unable to recognize the 4th stripe has been added to a 3-stripe Adidas logo) scoring an average of 17.05% accuracy in counting (e.g., counting stripes in an Adidas-like logo) across 7 diverse domains from animals, logos, chess, board games, optical illusions, to patterned grids. Removing image backgrounds nearly doubles accuracy (21.09 percentage points), revealing that contextual visual cues trigger these biased responses. Further analysis of VLMs' reasoning patterns shows that counting accuracy initially rises with thinking tokens, reaching ~40%, before declining with excessive reasoning. Our work presents an interesting failure mode in VLMs and a human-supervised automated framework for testing VLM biases. Code and data are available at: vlmsarebiased.github.io.

Dinh Phu Tran, Thao Do, Saad Wazir et al.

Transformer-based approaches have revolutionized image super-resolution by modeling long-range dependencies. However, the quadratic computational complexity of vanilla self-attention mechanisms poses significant challenges, often leading to compromises between efficiency and global context exploitation. Recent window-based attention methods mitigate this by localizing computations, but they often yield restricted receptive fields. To mitigate these limitations, we propose Selective Aggregation Transformer (SAT). This novel transformer efficiently captures long-range dependencies, leading to an enlarged model receptive field by selectively aggregating key-value matrices (reducing the number of tokens by 97\%) via our Density-driven Token Aggregation algorithm while maintaining the full resolution of the query matrix. This design significantly reduces computational costs, resulting in lower complexity and enabling scalable global interactions without compromising reconstruction fidelity. SAT identifies and represents each cluster with a single aggregation token, utilizing density and isolation metrics to ensure that critical high-frequency details are preserved. Experimental results demonstrate that SAT outperforms the state-of-the-art method PFT by up to 0.22dB, while the total number of FLOPs can be reduced by up to 27\%.

Dinh Phu Tran, Dao Duy Hung, Daeyoung Kim

Video super-resolution remains a major challenge in low-level vision tasks. To date, CNN- and Transformer-based methods have delivered impressive results. However, CNNs are limited by local receptive fields, while Transformers struggle with quadratic complexity, posing challenges for processing long sequences in VSR. Recently, Mamba has drawn attention for its long-sequence modeling, linear complexity, and large receptive fields. In this work, we propose VSRM, a novel \textbf{V}ideo \textbf{S}uper-\textbf{R}esolution framework that leverages the power of \textbf{M}amba. VSRM introduces Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to extract long-range spatio-temporal features and enhance receptive fields efficiently. To better align adjacent frames, we propose Deformable Cross-Mamba Alignment module. This module utilizes a deformable cross-mamba mechanism to make the compensation stage more dynamic and flexible, preventing feature distortions. Finally, we minimize the frequency domain gaps between reconstructed and ground-truth frames by proposing a simple yet effective Frequency Charbonnier-like loss that better preserves high-frequency content and enhances visual quality. Through extensive experiments, VSRM achieves state-of-the-art results on diverse benchmarks, establishing itself as a solid foundation for future research.

An Vo, Mohammad Reza Taesiri, Daeyoung Kim et al.

Large language models (LLMs) often exhibit strong biases, e.g, against women or in favor of the number 7. We investigate whether LLMs would be able to output less biased answers when allowed to observe their prior answers to the same question in a multi-turn conversation. To understand which types of questions invite more biased answers, we test LLMs on our proposed set of questions that span 9 topics and belong to three types: (1) Subjective; (2) Random; and (3) Objective. Interestingly, LLMs are able to "de-bias" themselves in a multi-turn conversation in response to questions that seek an Random, unbiased answer. Furthermore, we propose B-score, a novel metric that is effective in detecting biases to Subjective, Random, Easy, and Hard questions. On MMLU, HLE, and CSQA, leveraging B-score substantially improves the verification accuracy of LLM answers (i.e, accepting LLM correct answers and rejecting incorrect ones) compared to using verbalized confidence scores or the frequency of single-turn answers alone. Code and data are available at: https://b-score.github.io.

Jeongseok Kim, Jemin Lee, Yongin Kwon et al.

Mixed-precision quantization methods have been proposed to reduce model size while minimizing accuracy degradation. However, existing studies require retraining and do not consider the computational overhead and intermediate representations (IR) generated during the compilation process, limiting their application at the compiler level. This computational overhead refers to the runtime latency caused by frequent quantization and dequantization operations during inference. Performing these operations at the individual operator level causes significant runtime delays. To address these issues, we propose QuantuneV2, a compiler-based mixed-precision quantization method designed for practical embedded AI applications. QuantuneV2 performs inference only twice, once before quantization and once after quantization, and operates with a computational complexity of O(n) that increases linearly with the number of model parameters. We also made the sensitivity analysis more stable by using local metrics like weights, activation values, the Signal to Quantization Noise Ratio, and the Mean Squared Error. We also cut down on computational overhead by choosing the best IR and using operator fusion. Experimental results show that QuantuneV2 achieved up to a 10.28 percent improvement in accuracy and a 12.52 percent increase in speed compared to existing methods across five models: ResNet18v1, ResNet50v1, SqueezeNetv1, VGGNet, and MobileNetv2. This demonstrates that QuantuneV2 enhances model performance while maintaining computational efficiency, making it suitable for deployment in embedded AI environments.

Phuong Dam, Jihoon Jeong, Anh Tran et al.

This study discusses the critical issues of Virtual Try-On in contemporary e-commerce and the prospective metaverse, emphasizing the challenges of preserving intricate texture details and distinctive features of the target person and the clothes in various scenarios, such as clothing texture and identity characteristics like tattoos or accessories. In addition to the fidelity of the synthesized images, the efficiency of the synthesis process presents a significant hurdle. Various existing approaches are explored, highlighting the limitations and unresolved aspects, e.g., identity information omission, uncontrollable artifacts, and low synthesis speed. It then proposes a novel diffusion-based solution that addresses garment texture preservation and user identity retention during virtual try-on. The proposed network comprises two primary modules - a warping module aligning clothing with individual features and a try-on module refining the attire and generating missing parts integrated with a mask-aware post-processing technique ensuring the integrity of the individual's identity. It demonstrates impressive results, surpassing the state-of-the-art in speed by nearly 20 times during inference, with superior fidelity in qualitative assessments. Quantitative evaluations confirm comparable performance with the recent SOTA method on the VITON-HD and Dresscode datasets. We named our model Fast and Identity Preservation Virtual TryON (FIP-VITON).

Daeyoung Kim

Due to silence in early stages, lung cancer has been one of the most leading causes of mortality in cancer patients world-wide. Moreover, major symptoms of lung cancer are hard to differentiate with other respiratory disease symptoms such as COPD, further leading patients to overlook cancer progression in early stages. Thus, to enhance survival rates in lung cancer, early detection from consistent proactive respiratory system monitoring becomes crucial. One of the most prevalent and effective methods for lung cancer monitoring would be low-dose computed tomography(LDCT) chest scans, which led to remarkable enhancements in lung cancer detection or tumor classification tasks under rapid advancements and applications of computer vision based AI models such as EfficientNet or ResNet in image processing. However, though advanced CNN models under transfer learning or ViT based models led to high performing lung cancer detections, due to its intrinsic limitations in terms of correlation dependence and low interpretability due to complexity, expansions of deep learning models to lung cancer treatment analysis or causal intervention analysis simulations are still limited. Therefore, this research introduced LungCRCT: a latent causal representation learning based lung cancer analysis framework that retrieves causal representations of factors within the physical causal mechanism of lung cancer progression. With the use of advanced graph autoencoder based causal discovery algorithms with distance Correlation disentanglement and entropy-based image reconstruction refinement, LungCRCT not only enables causal intervention analysis for lung cancer treatments, but also leads to robust, yet extremely light downstream models in malignant tumor classification tasks with an AUC score of 93.91%.

Daeyoung Kim

As a representative optic degenerative condition, glaucoma has been a threat to millions due to its irreversibility and severe impact on human vision fields. Mainly characterized by dimmed and blurred visions, or peripheral vision loss, glaucoma is well known to occur due to damages in the optic nerve from increased intraocular pressure (IOP) or neovascularization within the retina. Traditionally, most glaucoma related works and clinical diagnosis focused on detecting these damages in the optic nerve by using patient data from perimetry tests, optic papilla inspections and tonometer-based IOP measurements. Recently, with advancements in computer vision AI models, such as VGG16 or Vision Transformers (ViT), AI-automatized glaucoma detection and optic cup segmentation based on retinal fundus images or OCT recently exhibited significant performance in aiding conventional diagnosis with high performance. However, current AI-driven glaucoma detection approaches still have significant room for improvement in terms of reliability, excessive parameter usage, possibility of spurious correlation within detection, and limitations in applications to intervention analysis or clinical simulations. Thus, this research introduced a novel causal representation driven glaucoma detection model: LightHCG, an extremely lightweight Convolutional VAE-based latent glaucoma representation model that can consider the true causality among glaucoma-related physical factors within the optic nerve region. Using HSIC-based latent space disentanglement and Graph Autoencoder based unsupervised causal representation learning, LightHCG not only exhibits higher performance in classifying glaucoma with 93~99% less weights, but also enhances the possibility of AI-driven intervention analysis, compared to existing advanced vision models such as InceptionV3, MobileNetV2 or VGG16.

Hyeontaek Hwang, Nguyen Dinh Son, Daeyoung Kim

Fine-tuning Multimodal Large Language Models (MLLMs) on task-specific data is an effective way to improve performance on downstream applications. However, such adaptation often leads to a degradation in generalization on pretrained tasks, a phenomenon known as Catastrophic Forgetting. Existing methods that aim to mitigate this issue either become ineffective when fine-tuning deeper layers of the language decoder or scale poorly with increasing model size. To address these limitations, we propose Model-Dowser, a novel sparse fine-tuning approach for MLLMs. Model-Dowser measures a principled importance score for each model parameter with respect to pretrained generalization (prior to downstream adaptation) by jointly considering weight magnitudes, input activations, and output sensitivities. During fine-tuning, Model-Dowser selectively preserves high-importance parameters and updates the remaining. Comprehensive experiments on two representative MLLMs, LLaVA and NVILA, demonstrate that Model-Dowser effectively mitigates catastrophic forgetting and consistently outperforms prior methods, while remaining resource-efficient and scalable to multi-billion-parameter models.

Daeyoung Kim

Age Related Macular Degeneration(AMD) has been one of the most leading causes of permanent vision impairment in ophthalmology. Though treatments, such as anti VEGF drugs or photodynamic therapies, were developed to slow down the degenerative process of AMD, there is still no specific cure to reverse vision loss caused by AMD. Thus, for AMD, detecting existence of risk factors of AMD or AMD itself within the patient retina in early stages is a crucial task to reduce the possibility of vision impairment. Apart from traditional approaches, deep learning based methods, especially attention mechanism based CNNs and GradCAM based XAI analysis on OCT scans, exhibited successful performance in distinguishing AMD retina from normal retinas, making it possible to use AI driven models to aid medical diagnosis and analysis by ophthalmologists regarding AMD. However, though having significant success, previous works mostly focused on prediction performance itself, not pathologies or underlying causal mechanisms of AMD, which can prohibit intervention analysis on specific factors or even lead to less reliable decisions. Thus, this paper introduces a novel causal AMD analysis model: GCVAMD, which incorporates a modified CausalVAE approach that can extract latent causal factors from only raw OCT images. By considering causality in AMD detection, GCVAMD enables causal inference such as treatment simulation or intervention analysis regarding major risk factors: drusen and neovascularization, while returning informative latent causal features that can enhance downstream tasks. Results show that through GCVAMD, drusen status and neovascularization status can be identified with AMD causal mechanisms in GCVAMD latent spaces, which can in turn be used for various tasks from AMD detection(classification) to intervention analysis.

Son Nguyen, Giang Nguyen, Hung Dao et al.

Key Information Extraction (KIE) underpins the understanding of visual documents (e.g., receipts and contracts) by extracting precise semantic content and accurately capturing spatial structure. Yet existing multimodal large language models (MLLMs) often perform poorly on dense documents and rely on vision tokenization approaches that scale with image size, leading to redundant computation and memory inefficiency. To address these challenges, we introduce VDInstruct, an MLLM that separates spatial region detection from semantic feature extraction. Central to our model is a content-aware tokenization strategy: rather than fragmenting the entire image uniformly, it generates tokens in proportion to document complexity, preserving critical structure while eliminating wasted tokens. Leveraging a three-stage training paradigm, our model achieves state-of-the-art (SOTA) results on KIE benchmarks, matching or exceeding the accuracy of leading approaches while reducing the number of image tokens by roughly 3.6x. In zero-shot evaluations, VDInstruct surpasses strong baselines-such as DocOwl 1.5-by +5.5 F1 points, highlighting its robustness to unseen documents. These findings show that content-aware tokenization combined with explicit layout modeling offers a promising direction forward for document understanding. Data, source code, and model weights will be made publicly available.

Seongsu Bae, Daeyoung Kim, Jiho Kim et al.

An intelligent machine that can answer human questions based on electronic health records (EHR-QA) has a great practical value, such as supporting clinical decisions, managing hospital administration, and medical chatbots. Previous table-based QA studies focusing on translating natural questions into table queries (NLQ2SQL), however, suffer from the unique nature of EHR data due to complex and specialized medical terminology, hence increased decoding difficulty. In this paper, we design UniQA, a unified encoder-decoder architecture for EHR-QA where natural language questions are converted to queries such as SQL or SPARQL. We also propose input masking (IM), a simple and effective method to cope with complex medical terms and various typos and better learn the SQL/SPARQL syntax. Combining the unified architecture with an effective auxiliary training objective, UniQA demonstrated a significant performance improvement against the previous state-of-the-art model for MIMICSQL* (14.2% gain), the most complex NLQ2SQL dataset in the EHR domain, and its typo-ridden versions (approximately 28.8% gain). In addition, we confirmed consistent results for the graph-based EHR-QA dataset, MIMICSPARQL*.

Trung Q. Tran, Mingu Kang, Daeyoung Kim

Semi-supervised learning (SSL) has played an important role in leveraging unlabeled data when labeled data is limited. One of the most successful SSL approaches is based on consistency regularization, which encourages the model to produce unchanged with perturbed input. However, there has been less attention spent on inputs that have the same label. Motivated by the observation that the inputs having the same label should have the similar model outputs, we propose a novel method, RankingMatch, that considers not only the perturbed inputs but also the similarity among the inputs having the same label. We especially introduce a new objective function, dubbed BatchMean Triplet loss, which has the advantage of computational efficiency while taking into account all input samples. Our RankingMatch achieves state-of-the-art performance across many standard SSL benchmarks with a variety of labeled data amounts, including 95.13% accuracy on CIFAR-10 with 250 labels, 77.65% accuracy on CIFAR-100 with 10000 labels, 97.76% accuracy on SVHN with 250 labels, and 97.77% accuracy on SVHN with 1000 labels. We also perform an ablation study to prove the efficacy of the proposed BatchMean Triplet loss against existing versions of Triplet loss.

Giang Nguyen, Daeyoung Kim, Anh Nguyen

Explaining the decisions of an Artificial Intelligence (AI) model is increasingly critical in many real-world, high-stake applications. Hundreds of papers have either proposed new feature attribution methods, discussed or harnessed these tools in their work. However, despite humans being the target end-users, most attribution methods were only evaluated on proxy automatic-evaluation metrics (Zhang et al. 2018; Zhou et al. 2016; Petsiuk et al. 2018). In this paper, we conduct the first user study to measure attribution map effectiveness in assisting humans in ImageNet classification and Stanford Dogs fine-grained classification, and when an image is natural or adversarial (i.e., contains adversarial perturbations). Overall, feature attribution is surprisingly not more effective than showing humans nearest training-set examples. On a harder task of fine-grained dog categorization, presenting attribution maps to humans does not help, but instead hurts the performance of human-AI teams compared to AI alone. Importantly, we found automatic attribution-map evaluation measures to correlate poorly with the actual human-AI team performance. Our findings encourage the community to rigorously test their methods on the downstream human-in-the-loop applications and to rethink the existing evaluation metrics.

Thao Do, Daeyoung Kim

There are many real-life use cases such as barcode scanning or billboard reading where people need to detect objects and read the object contents. Commonly existing methods are first trying to localize object regions, then determine layout and lastly classify content units. However, for simple fixed structured objects like license plates, this approach becomes overkill and lengthy to run. This work aims to solve this detect-and-read problem in a lightweight way by integrating multi-digit recognition into a one-stage object detection model. Our unified method not only eliminates the duplication in feature extraction (one for localizing, one again for classifying) but also provides useful contextual information around object regions for classification. Additionally, our choice of backbones and modifications in architecture, loss function, data augmentation and training make the method robust, efficient and speedy. Secondly, we made a public benchmark dataset of diverse real-life 1D barcodes for a reliable evaluation, which we collected, annotated and checked carefully. Eventually, experimental results prove the method's efficiency on the barcode problem by outperforming industrial tools in both detecting and decoding rates with a real-time fps at a VGA-similar resolution. It also did a great job expectedly on the license-plate recognition task (on the AOLP dataset) by outperforming the current state-of-the-art method significantly in terms of recognition rate and inference time.

Mingu Kang, Trung Quang Tran, Seungju Cho et al.

Adversarial attack is aimed at fooling the target classifier with imperceptible perturbation. Adversarial examples, which are carefully crafted with a malicious purpose, can lead to erroneous predictions, resulting in catastrophic accidents. To mitigate the effects of adversarial attacks, we propose a novel purification model called CAP-GAN. CAP-GAN takes account of the idea of pixel-level and feature-level consistency to achieve reasonable purification under cycle-consistent learning. Specifically, we utilize the guided attention module and knowledge distillation to convey meaningful information to the purification model. Once a model is fully trained, inputs would be projected into the purification model and transformed into clean-like images. We vary the capacity of the adversary to argue the robustness against various types of attack strategies. On the CIFAR-10 dataset, CAP-GAN outperforms other pre-processing based defenses under both black-box and white-box settings.

Daeyoung Kim, Esteban G. Tabak

A novel algorithm is proposed to solve the sample-based optimal transport problem. An adversarial formulation of the push-forward condition uses a test function built as a convolution between an adaptive kernel and an evolving probability distribution $ν$ over a latent variable $b$. Approximating this convolution by its simulation over evolving samples $b^i(t)$ of $ν$, the parameterization of the test function reduces to determining the flow of these samples. This flow, discretized over discrete time steps $t_n$, is built from the composition of elementary maps. The optimal transport also follows a flow that, by duality, must follow the gradient of the test function. The representation of the test function as the Monte Carlo simulation of a distribution makes the algorithm robust to dimensionality, and its evolution under a memory-less flow produces rich, complex maps from simple parametric transformations. The algorithm is illustrated with numerical examples.

Giang Nguyen, Shuan Chen, Tae Joon Jun et al.

Explaining the behaviors of deep neural networks, usually considered as black boxes, is critical especially when they are now being adopted over diverse aspects of human life. Taking the advantages of interpretable machine learning (interpretable ML), this paper proposes a novel tool called Catastrophic Forgetting Dissector (or CFD) to explain catastrophic forgetting in continual learning settings. We also introduce a new method called Critical Freezing based on the observations of our tool. Experiments on ResNet articulate how catastrophic forgetting happens, particularly showing which components of this famous network are forgetting. Our new continual learning algorithm defeats various recent techniques by a significant margin, proving the capability of the investigation. Critical freezing not only attacks catastrophic forgetting but also exposes explainability.

Thao Do, Yalew Tolcha, Tae Joon Jun et al.

Barcodes are ubiquitous and have been used in most of critical daily activities for decades. However, most of traditional decoders require well-founded barcode under a relatively standard condition. While wilder conditioned barcodes such as underexposed, occluded, blurry, wrinkled and rotated are commonly captured in reality, those traditional decoders show weakness of recognizing. Several works attempted to solve those challenging barcodes, but many limitations still exist. This work aims to solve the decoding problem using deep convolutional neural network with the possibility of running on portable devices. Firstly, we proposed a special modification of inference based on the feature of having checksum and test-time augmentation, named as Smart Inference (SI) in prediction phase of a trained model. SI considerably boosts accuracy and reduces the false prediction for trained models. Secondly, we have created a large practical evaluation dataset of real captured 1D barcode under various challenging conditions to test our methods vigorously, which is publicly available for other researchers. The experiments' results demonstrated the SI effectiveness with the highest accuracy of 95.85% which outperformed many existing decoders on the evaluation set. Finally, we successfully minimized the best model by knowledge distillation to a shallow model which is shown to have high accuracy (90.85%) with good inference speed of 34.2 ms per image on a real edge device.

Seungju Cho, Tae Joon Jun, Mingu Kang et al.

Nowadays the deep learning technology is growing faster and shows dramatic performance in computer vision areas. However, it turns out a deep learning based model is highly vulnerable to some small perturbation called an adversarial attack. It can easily fool the deep learning model by adding small perturbations. On the other hand, tensor decomposition method widely uses for compressing the tensor data, including data matrix, image, etc. In this paper, we suggest combining tensor decomposition for defending the model against adversarial example. We verify this idea is simple and effective to resist adversarial attack. In addition, this method rarely degrades the original performance of clean data. We experiment on MNIST, CIFAR10 and ImageNet data and show our method robust on state-of-the-art attack methods.

Hyungrok Ham, Tae Joon Jun, Daeyoung Kim

We propose Unbalanced GANs, which pre-trains the generator of the generative adversarial network (GAN) using variational autoencoder (VAE). We guarantee the stable training of the generator by preventing the faster convergence of the discriminator at early epochs. Furthermore, we balance between the generator and the discriminator at early epochs and thus maintain the stabilized training of GANs. We apply Unbalanced GANs to well known public datasets and find that Unbalanced GANs reduce mode collapses. We also show that Unbalanced GANs outperform ordinary GANs in terms of stabilized learning, faster convergence and better image quality at early epochs.

Giang Nguyen, Shuan Chen, Thao Do et al.

Interpreting the behaviors of Deep Neural Networks (usually considered as a black box) is critical especially when they are now being widely adopted over diverse aspects of human life. Taking the advancements from Explainable Artificial Intelligent, this paper proposes a novel technique called Auto DeepVis to dissect catastrophic forgetting in continual learning. A new method to deal with catastrophic forgetting named critical freezing is also introduced upon investigating the dilemma by Auto DeepVis. Experiments on a captioning model meticulously present how catastrophic forgetting happens, particularly showing which components are forgetting or changing. The effectiveness of our technique is then assessed; and more precisely, critical freezing claims the best performance on both previous and coming tasks over baselines, proving the capability of the investigation. Our techniques could not only be supplementary to existing solutions for completely eradicating catastrophic forgetting for life-long learning but also explainable.

Giang Nguyen, Tae Joon Jun, Trung Tran et al.

While advanced image captioning systems are increasingly describing images coherently and exactly, recent progress in continual learning allows deep learning models to avoid catastrophic forgetting. However, the domain where image captioning working with continual learning has not yet been explored. We define the task in which we consolidate continual learning and image captioning as continual image captioning. In this work, we propose ContCap, a framework generating captions over a series of new tasks coming, seamlessly integrating continual learning into image captioning besides addressing catastrophic forgetting. After proving forgetting in image captioning, we propose various techniques to overcome the forgetting dilemma by taking a simple fine-tuning schema as the baseline. We split MS-COCO 2014 dataset to perform experiments in class-incremental settings without revisiting dataset of previously provided tasks. Experiments show remarkable improvements in the performance on the old tasks while the figures for the new surprisingly surpass fine-tuning. Our framework also offers a scalable solution for continual image or video captioning.

Seungju Cho, Tae Joon Jun, Byungsoo Oh et al.

Nowadays, Deep learning techniques show dramatic performance on computer vision area, and they even outperform human. But it is also vulnerable to some small perturbation called an adversarial attack. This is a problem combined with the safety of artificial intelligence, which has recently been studied a lot. These attacks have shown that they can fool models of image classification, semantic segmentation, and object detection. We point out this attack can be protected by denoise autoencoder, which is used for denoising the perturbation and restoring the original images. We experiment with various noise distributions and verify the effect of denoise autoencoder against adversarial attack in semantic segmentation.

Tae Joon Jun, Youngsub Eom, Dohyeun Kim et al.

In this paper, we proposed Transferable Ranking Convolutional Neural Network (TRk-CNN) that can be effectively applied when the classes of images to be classified show a high correlation with each other. The multi-class classification method based on the softmax function, which is generally used, is not effective in this case because the inter-class relationship is ignored. Although there is a Ranking-CNN that takes into account the ordinal classes, it cannot reflect the inter-class relationship to the final prediction. TRk-CNN, on the other hand, combines the weights of the primitive classification model to reflect the inter-class information to the final classification phase. We evaluated TRk-CNN in glaucoma image dataset that was labeled into three classes: normal, glaucoma suspect, and glaucoma eyes. Based on the literature we surveyed, this study is the first to classify three status of glaucoma fundus image dataset into three different classes. We compared the evaluation results of TRk-CNN with Ranking-CNN (Rk-CNN) and multi-class CNN (MC-CNN) using the DenseNet as the backbone CNN model. As a result, TRk-CNN achieved an average accuracy of 92.96%, specificity of 93.33%, sensitivity for glaucoma suspect of 95.12% and sensitivity for glaucoma of 93.98%. Based on average accuracy, TRk-CNN is 8.04% and 9.54% higher than Rk-CNN and MC-CNN and surprisingly 26.83% higher for sensitivity for suspicious than multi-class CNN. Our TRk-CNN is expected to be effectively applied to the medical image classification problem where the disease state is continuous and increases in the positive class direction.

Tae Joon Jun, Jihoon Kweon, Young-Hak Kim et al.

In this paper, we proposed T-Net containing a small encoder-decoder inside the encoder-decoder structure (EDiED). T-Net overcomes the limitation that U-Net can only have a single set of the concatenate layer between encoder and decoder block. To be more precise, the U-Net symmetrically forms the concatenate layers, so the low-level feature of the encoder is connected to the latter part of the decoder, and the high-level feature is connected to the beginning of the decoder. T-Net arranges the pooling and up-sampling appropriately during the encoder process, and likewise during the decoding process so that feature-maps of various sizes are obtained in a single block. As a result, all features from the low-level to the high-level extracted from the encoder are delivered from the beginning of the decoder to predict a more accurate mask. We evaluated T-Net for the problem of segmenting three main vessels in coronary angiography images. The experiment consisted of a comparison of U-Net and T-Nets under the same conditions, and an optimized T-Net for the main vessel segmentation. As a result, T-Net recorded a Dice Similarity Coefficient score (DSC) of 0.815, 0.095 higher than that of U-Net, and the optimized T-Net recorded a DSC of 0.890 which was 0.170 higher than that of U-Net. In addition, we visualized the weight activation of the convolutional layer of T-Net and U-Net to show that T-Net actually predicts the mask from earlier decoders. Therefore, we expect that T-Net can be effectively applied to other similar medical image segmentation problems.

Dohyeun Kim, Tae Joon Jun, Daeyoung Kim et al.

Solving the classification problem, unbalanced number of dataset among the classes often causes performance degradation. Especially when some classes dominate the other classes with its large number of datasets, trained model shows low performance in identifying the dominated classes. This is common case when it comes to medical dataset. Because the case with a serious degree is not quite usual, there are imbalance in number of dataset between severe case and normal cases of diseases. Also, there is difficulty in precisely identifying grade of medical data because of vagueness between them. To solve these problems, we propose new architecture of convolutional neural network named Tournament based Ranking CNN which shows remarkable performance gain in identifying dominated classes while trading off very small accuracy loss in dominating classes. Our Approach complemented problems that occur when method of Ranking CNN that aggregates outputs of multiple binary neural network models is applied to medical data. By having tournament structure in aggregating method and using very deep pretrained binary models, our proposed model recorded 68.36% of exact match accuracy, while Ranking CNN recorded 53.40%, pretrained Resnet recorded 56.12% and CNN with linear regression recorded 57.48%. As a result, our proposed method is applied efficiently to cataract grading which have ordinal labels with imbalanced number of data among classes, also can be applied further to medical problems which have similar features to cataract and similar dataset configuration.