Hong Liu

Xingbin Liu, Huafeng Kuang, Hong Liu et al.

Deep neural networks have been applied in many computer vision tasks and achieved state-of-the-art performance. However, misclassification will occur when DNN predicts adversarial examples which add human-imperceptible adversarial noise to natural examples. This limits the application of DNN in security-critical fields. To alleviate this problem, we first conducted an empirical analysis of the latent features of both adversarial and natural examples and found the similarity matrix of natural examples is more compact than those of adversarial examples. Motivated by this observation, we propose \textbf{L}atent \textbf{F}eature \textbf{R}elation \textbf{C}onsistency (\textbf{LFRC}), which constrains the relation of adversarial examples in latent space to be consistent with the natural examples. Importantly, our LFRC is orthogonal to the previous method and can be easily combined with them to achieve further improvement. To demonstrate the effectiveness of LFRC, we conduct extensive experiments using different neural networks on benchmark datasets. For instance, LFRC can bring 0.78\% further improvement compared to AT, and 1.09\% improvement compared to TRADES, against AutoAttack on CIFAR10. Code is available at https://github.com/liuxingbin/LFRC.

Coen de Vente, Koenraad A. Vermeer, Nicolas Jaccard et al.

The early detection of glaucoma is essential in preventing visual impairment. Artificial intelligence (AI) can be used to analyze color fundus photographs (CFPs) in a cost-effective manner, making glaucoma screening more accessible. While AI models for glaucoma screening from CFPs have shown promising results in laboratory settings, their performance decreases significantly in real-world scenarios due to the presence of out-of-distribution and low-quality images. To address this issue, we propose the Artificial Intelligence for Robust Glaucoma Screening (AIROGS) challenge. This challenge includes a large dataset of around 113,000 images from about 60,000 patients and 500 different screening centers, and encourages the development of algorithms that are robust to ungradable and unexpected input data. We evaluated solutions from 14 teams in this paper, and found that the best teams performed similarly to a set of 20 expert ophthalmologists and optometrists. The highest-scoring team achieved an area under the receiver operating characteristic curve of 0.99 (95% CI: 0.98-0.99) for detecting ungradable images on-the-fly. Additionally, many of the algorithms showed robust performance when tested on three other publicly available datasets. These results demonstrate the feasibility of robust AI-enabled glaucoma screening.

Xuxin Chen, Ke Zhang, Neman Abdoli et al.

Deep convolutional neural networks (CNNs) have been widely used in various medical imaging tasks. However, due to the intrinsic locality of convolution operation, CNNs generally cannot model long-range dependencies well, which are important for accurately identifying or mapping corresponding breast lesion features computed from unregistered multiple mammograms. This motivates us to leverage the architecture of Multi-view Vision Transformers to capture long-range relationships of multiple mammograms from the same patient in one examination. For this purpose, we employ local Transformer blocks to separately learn patch relationships within four mammograms acquired from two-view (CC/MLO) of two-side (right/left) breasts. The outputs from different views and sides are concatenated and fed into global Transformer blocks, to jointly learn patch relationships between four images representing two different views of the left and right breasts. To evaluate the proposed model, we retrospectively assembled a dataset involving 949 sets of mammograms, which include 470 malignant cases and 479 normal or benign cases. We trained and evaluated the model using a five-fold cross-validation method. Without any arduous preprocessing steps (e.g., optimal window cropping, chest wall or pectoral muscle removal, two-view image registration, etc.), our four-image (two-view-two-side) Transformer-based model achieves case classification performance with an area under ROC curve (AUC = 0.818), which significantly outperforms AUC = 0.784 achieved by the state-of-the-art multi-view CNNs (p = 0.009). It also outperforms two one-view-two-side models that achieve AUC of 0.724 (CC view) and 0.769 (MLO view), respectively. The study demonstrates the potential of using Transformers to develop high-performing computer-aided diagnosis schemes that combine four mammograms.

Jinghan Sun, Dong Wei, Zhe Xu et al.

Chest X-ray (CXR) anatomical abnormality detection aims at localizing and characterising cardiopulmonary radiological findings in the radiographs, which can expedite clinical workflow and reduce observational oversights. Most existing methods attempted this task in either fully supervised settings which demanded costly mass per-abnormality annotations, or weakly supervised settings which still lagged badly behind fully supervised methods in performance. In this work, we propose a co-evolutionary image and report distillation (CEIRD) framework, which approaches semi-supervised abnormality detection in CXR by grounding the visual detection results with text-classified abnormalities from paired radiology reports, and vice versa. Concretely, based on the classical teacher-student pseudo label distillation (TSD) paradigm, we additionally introduce an auxiliary report classification model, whose prediction is used for report-guided pseudo detection label refinement (RPDLR) in the primary vision detection task. Inversely, we also use the prediction of the vision detection model for abnormality-guided pseudo classification label refinement (APCLR) in the auxiliary report classification task, and propose a co-evolution strategy where the vision and report models mutually promote each other with RPDLR and APCLR performed alternatively. To this end, we effectively incorporate the weak supervision by reports into the semi-supervised TSD pipeline. Besides the cross-modal pseudo label refinement, we further propose an intra-image-modal self-adaptive non-maximum suppression, where the pseudo detection labels generated by the teacher vision model are dynamically rectified by high-confidence predictions by the student. Experimental results on the public MIMIC-CXR benchmark demonstrate CEIRD's superior performance to several up-to-date weakly and semi-supervised methods.

Hong Liu, Dong Wei, Donghuan Lu et al.

Automated surface segmentation of retinal layer is important and challenging in analyzing optical coherence tomography (OCT). Recently, many deep learning based methods have been developed for this task and yield remarkable performance. However, due to large spatial gap and potential mismatch between the B-scans of OCT data, all of them are based on 2D segmentation of individual B-scans, which may loss the continuity information across the B-scans. In addition, 3D surface of the retina layers can provide more diagnostic information, which is crucial in quantitative image analysis. In this study, a novel framework based on hybrid 2D-3D convolutional neural networks (CNNs) is proposed to obtain continuous 3D retinal layer surfaces from OCT. The 2D features of individual B-scans are extracted by an encoder consisting of 2D convolutions. These 2D features are then used to produce the alignment displacement field and layer segmentation by two 3D decoders, which are coupled via a spatial transformer module. The entire framework is trained end-to-end. To the best of our knowledge, this is the first study that attempts 3D retinal layer segmentation in volumetric OCT images based on CNNs. Experiments on a publicly available dataset show that our framework achieves superior results to state-of-the-art 2D methods in terms of both layer segmentation accuracy and cross-B-scan 3D continuity, thus offering more clinical values than previous works.

Zhifang Guo, Jianguo Mao, Rui Tao et al.

Text-based audio generation models have limitations as they cannot encompass all the information in audio, leading to restricted controllability when relying solely on text. To address this issue, we propose a novel model that enhances the controllability of existing pre-trained text-to-audio models by incorporating additional conditions including content (timestamp) and style (pitch contour and energy contour) as supplements to the text. This approach achieves fine-grained control over the temporal order, pitch, and energy of generated audio. To preserve the diversity of generation, we employ a trainable control condition encoder that is enhanced by a large language model and a trainable Fusion-Net to encode and fuse the additional conditions while keeping the weights of the pre-trained text-to-audio model frozen. Due to the lack of suitable datasets and evaluation metrics, we consolidate existing datasets into a new dataset comprising the audio and corresponding conditions and use a series of evaluation metrics to evaluate the controllability performance. Experimental results demonstrate that our model successfully achieves fine-grained control to accomplish controllable audio generation. Audio samples and our dataset are publicly available at https://conditionaudiogen.github.io/conditionaudiogen/

Hong Liu, Liren Shan, Han Bao et al.

Federated learning is often used in environments with many unverified participants. Therefore, federated learning under adversarial attacks receives significant attention. This paper proposes an algorithmic framework for list-decodable federated learning, where a central server maintains a list of models, with at least one guaranteed to perform well. The framework has no strict restriction on the fraction of honest workers, extending the applicability of Byzantine federated learning to the scenario with more than half adversaries. Under proper assumptions on the loss function, we prove a convergence theorem for our method. Experimental results, including image classification tasks with both convex and non-convex losses, demonstrate that the proposed algorithm can withstand the malicious majority under various attacks.

Jinghan Sun, Dong Wei, Zhe Xu et al.

Anatomical abnormality detection and report generation of chest X-ray (CXR) are two essential tasks in clinical practice. The former aims at localizing and characterizing cardiopulmonary radiological findings in CXRs, while the latter summarizes the findings in a detailed report for further diagnosis and treatment. Existing methods often focused on either task separately, ignoring their correlation. This work proposes a co-evolutionary abnormality detection and report generation (CoE-DG) framework. The framework utilizes both fully labeled (with bounding box annotations and clinical reports) and weakly labeled (with reports only) data to achieve mutual promotion between the abnormality detection and report generation tasks. Specifically, we introduce a bi-directional information interaction strategy with generator-guided information propagation (GIP) and detector-guided information propagation (DIP). For semi-supervised abnormality detection, GIP takes the informative feature extracted by the generator as an auxiliary input to the detector and uses the generator's prediction to refine the detector's pseudo labels. We further propose an intra-image-modal self-adaptive non-maximum suppression module (SA-NMS). This module dynamically rectifies pseudo detection labels generated by the teacher detection model with high-confidence predictions by the student.Inversely, for report generation, DIP takes the abnormalities' categories and locations predicted by the detector as input and guidance for the generator to improve the generated reports.

Xuxin Chen, Ximin Wang, Ke Zhang et al.

This study aims to develop a novel computer-aided diagnosis (CAD) scheme for mammographic breast mass classification using semi-supervised learning. Although supervised deep learning has achieved huge success across various medical image analysis tasks, its success relies on large amounts of high-quality annotations, which can be challenging to acquire in practice. To overcome this limitation, we propose employing a semi-supervised method, i.e., virtual adversarial training (VAT), to leverage and learn useful information underlying in unlabeled data for better classification of breast masses. Accordingly, our VAT-based models have two types of losses, namely supervised and virtual adversarial losses. The former loss acts as in supervised classification, while the latter loss aims at enhancing model robustness against virtual adversarial perturbation, thus improving model generalizability. To evaluate the performance of our VAT-based CAD scheme, we retrospectively assembled a total of 1024 breast mass images, with equal number of benign and malignant masses. A large CNN and a small CNN were used in this investigation, and both were trained with and without the adversarial loss. When the labeled ratios were 40% and 80%, VAT-based CNNs delivered the highest classification accuracy of 0.740 and 0.760, respectively. The experimental results suggest that the VAT-based CAD scheme can effectively utilize meaningful knowledge from unlabeled data to better classify mammographic breast mass images.

Hong Liu, Jeff Z. HaoChen, Adrien Gaidon et al.

Self-supervised learning (SSL) is a scalable way to learn general visual representations since it learns without labels. However, large-scale unlabeled datasets in the wild often have long-tailed label distributions, where we know little about the behavior of SSL. In this work, we systematically investigate self-supervised learning under dataset imbalance. First, we find out via extensive experiments that off-the-shelf self-supervised representations are already more robust to class imbalance than supervised representations. The performance gap between balanced and imbalanced pre-training with SSL is significantly smaller than the gap with supervised learning, across sample sizes, for both in-domain and, especially, out-of-domain evaluation. Second, towards understanding the robustness of SSL, we hypothesize that SSL learns richer features from frequent data: it may learn label-irrelevant-but-transferable features that help classify the rare classes and downstream tasks. In contrast, supervised learning has no incentive to learn features irrelevant to the labels from frequent examples. We validate this hypothesis with semi-synthetic experiments and theoretical analyses on a simplified setting. Third, inspired by the theoretical insights, we devise a re-weighted regularization technique that consistently improves the SSL representation quality on imbalanced datasets with several evaluation criteria, closing the small gap between balanced and imbalanced datasets with the same number of examples.

Zhirong Ye, Xiangdong Wang, Hong Liu et al.

Sound event detection (SED) has gained increasing attention with its wide application in surveillance, video indexing, etc. Existing models in SED mainly generate frame-level prediction, converting it into a sequence multi-label classification problem. A critical issue with the frame-based model is that it pursues the best frame-level prediction rather than the best event-level prediction. Besides, it needs post-processing and cannot be trained in an end-to-end way. This paper firstly presents the one-dimensional Detection Transformer (1D-DETR), inspired by Detection Transformer for image object detection. Furthermore, given the characteristics of SED, the audio query branch and a one-to-many matching strategy for fine-tuning the model are added to 1D-DETR to form Sound Event Detection Transformer (SEDT). To our knowledge, SEDT is the first event-based and end-to-end SED model. Experiments are conducted on the URBAN-SED dataset and the DCASE2019 Task4 dataset, and both show that SEDT can achieve competitive performance.

Morteza Heidari, Sivaramakrishnan Lakshmivarahan, Seyedehnafiseh Mirniaharikandehei et al.

Machine learning is widely used in developing computer-aided diagnosis (CAD) schemes of medical images. However, CAD usually computes large number of image features from the targeted regions, which creates a challenge of how to identify a small and optimal feature vector to build robust machine learning models. In this study, we investigate feasibility of applying a random projection algorithm to build an optimal feature vector from the initially CAD-generated large feature pool and improve performance of machine learning model. We assemble a retrospective dataset involving 1,487 cases of mammograms in which 644 cases have confirmed malignant mass lesions and 843 have benign lesions. A CAD scheme is first applied to segment mass regions and initially compute 181 features. Then, support vector machine (SVM) models embedded with several feature dimensionality reduction methods are built to predict likelihood of lesions being malignant. All SVM models are trained and tested using a leave-one-case-out cross-validation method. SVM generates a likelihood score of each segmented mass region depicting on one-view mammogram. By fusion of two scores of the same mass depicting on two-view mammograms, a case-based likelihood score is also evaluated. Comparing with the principle component analyses, nonnegative matrix factorization, and Chi-squared methods, SVM embedded with the random projection algorithm yielded a significantly higher case-based lesion classification performance with the area under ROC curve of 0.84+0.01 (p<0.02). The study demonstrates that the random project algorithm is a promising method to generate optimal feature vectors to help improve performance of machine learning models of medical images.

Yong Ruan, Xiangdong Wang, Hong Liu et al.

The dominant automatic lexical stress detection method is to split the utterance into syllable segments using phoneme sequence and their time-aligned boundaries. Then we extract features from syllable to use classification method to classify the lexical stress. However, we can't get very accurate time boundaries of each phoneme and we have to design some features in the syllable segments to classify the lexical stress. Therefore, we propose a end-to-end approach using sequence to sequence model of transformer to estimate lexical stress. For this, we train transformer model using feature sequence of audio and their phoneme sequence with lexical stress marks. During the recognition process, the recognized phoneme sequence is restricted according to the original standard phoneme sequence without lexical stress marks, but the lexical stress mark of each phoneme is not limited. We train the model in different subset of Librispeech and do lexical stress recognition in TIMIT and L2-ARCTIC dataset. For all subsets, the end-to-end model will perform better than the syllable segments classification method. Our method can achieve a 6.36% phoneme error rate on the TIMIT dataset, which exceeds the 7.2% error rate in other studies.

Jie Hu, Rongrong Ji, Hong Liu et al.

Most existing visual search systems are deployed based upon fixed kinds of visual features, which prohibits the feature reusing across different systems or when upgrading systems with a new type of feature. Such a setting is obviously inflexible and time/memory consuming, which is indeed mendable if visual features can be "translated" across systems. In this paper, we make the first attempt towards visual feature translation to break through the barrier of using features across different visual search systems. To this end, we propose a Hybrid Auto-Encoder (HAE) to translate visual features, which learns a mapping by minimizing the translation and reconstruction errors. Based upon HAE, an Undirected Affinity Measurement (UAM) is further designed to quantify the affinity among different types of visual features. Extensive experiments have been conducted on several public datasets with sixteen different types of widely-used features in visual search systems. Quantitative results show the encouraging possibilities of feature translation. For the first time, the affinity among widely-used features like SIFT and DELF is reported.

Jie Li, Rongrong Ji, Hong Liu et al.

Universal adversarial perturbations (UAPs), a.k.a. input-agnostic perturbations, has been proved to exist and be able to fool cutting-edge deep learning models on most of the data samples. Existing UAP methods mainly focus on attacking image classification models. Nevertheless, little attention has been paid to attacking image retrieval systems. In this paper, we make the first attempt in attacking image retrieval systems. Concretely, image retrieval attack is to make the retrieval system return irrelevant images to the query at the top ranking list. It plays an important role to corrupt the neighbourhood relationships among features in image retrieval attack. To this end, we propose a novel method to generate retrieval-against UAP to break the neighbourhood relationships of image features via degrading the corresponding ranking metric. To expand the attack method to scenarios with varying input sizes or untouchable network parameters, a multi-scale random resizing scheme and a ranking distillation strategy are proposed. We evaluate the proposed method on four widely-used image retrieval datasets, and report a significant performance drop in terms of different metrics, such as mAP and mP@10. Finally, we test our attack methods on the real-world visual search engine, i.e., Google Images, which demonstrates the practical potentials of our methods.

Dan Xu, Wei Wang, Hao Tang et al.

Recent works have shown the benefit of integrating Conditional Random Fields (CRFs) models into deep architectures for improving pixel-level prediction tasks. Following this line of research, in this paper we introduce a novel approach for monocular depth estimation. Similarly to previous works, our method employs a continuous CRF to fuse multi-scale information derived from different layers of a front-end Convolutional Neural Network (CNN). Differently from past works, our approach benefits from a structured attention model which automatically regulates the amount of information transferred between corresponding features at different scales. Importantly, the proposed attention model is seamlessly integrated into the CRF, allowing end-to-end training of the entire architecture. Our extensive experimental evaluation demonstrates the effectiveness of the proposed method which is competitive with previous methods on the KITTI benchmark and outperforms the state of the art on the NYU Depth V2 dataset.

Mengyuan Liu, Hong Liu, Chen Chen

3D action recognition has broad applications in human-computer interaction and intelligent surveillance. However, recognizing similar actions remains challenging since previous literature fails to capture motion and shape cues effectively from noisy depth data. In this paper, we propose a novel two-layer Bag-of-Visual-Words (BoVW) model, which suppresses the noise disturbances and jointly encodes both motion and shape cues. First, background clutter is removed by a background modeling method that is designed for depth data. Then, motion and shape cues are jointly used to generate robust and distinctive spatial-temporal interest points (STIPs): motion-based STIPs and shape-based STIPs. In the first layer of our model, a multi-scale 3D local steering kernel (M3DLSK) descriptor is proposed to describe local appearances of cuboids around motion-based STIPs. In the second layer, a spatial-temporal vector (STV) descriptor is proposed to describe the spatial-temporal distributions of shape-based STIPs. Using the Bag-of-Visual-Words (BoVW) model, motion and shape cues are combined to form a fused action representation. Our model performs favorably compared with common STIP detection and description methods. Thorough experiments verify that our model is effective in distinguishing similar actions and robust to background clutter, partial occlusions and pepper noise.

Hong Liu, Juanhui Tu, Mengyuan Liu

It remains a challenge to efficiently extract spatialtemporal information from skeleton sequences for 3D human action recognition. Although most recent action recognition methods are based on Recurrent Neural Networks which present outstanding performance, one of the shortcomings of these methods is the tendency to overemphasize the temporal information. Since 3D convolutional neural network(3D CNN) is a powerful tool to simultaneously learn features from both spatial and temporal dimensions through capturing the correlations between three dimensional signals, this paper proposes a novel two-stream model using 3D CNN. To our best knowledge, this is the first application of 3D CNN in skeleton-based action recognition. Our method consists of three stages. First, skeleton joints are mapped into a 3D coordinate space and then encoding the spatial and temporal information, respectively. Second, 3D CNN models are seperately adopted to extract deep features from two streams. Third, to enhance the ability of deep features to capture global relationships, we extend every stream into multitemporal version. Extensive experiments on the SmartHome dataset and the large-scale NTU RGB-D dataset demonstrate that our method outperforms most of RNN-based methods, which verify the complementary property between spatial and temporal information and the robustness to noise.