Yuli Wu

Reza Azad, Moein Heidari, Yuli Wu et al.

Currently, convolutional neural networks (CNN) (e.g., U-Net) have become the de facto standard and attained immense success in medical image segmentation. However, as a downside, CNN based methods are a double-edged sword as they fail to build long-range dependencies and global context connections due to the limited receptive field that stems from the intrinsic characteristics of the convolution operation. Hence, recent articles have exploited Transformer variants for medical image segmentation tasks which open up great opportunities due to their innate capability of capturing long-range correlations through the attention mechanism. Although being feasibly designed, most of the cohort studies incur prohibitive performance in capturing local information, thereby resulting in less lucidness of boundary areas. In this paper, we propose a contextual attention network to tackle the aforementioned limitations. The proposed method uses the strength of the Transformer module to model the long-range contextual dependency. Simultaneously, it utilizes the CNN encoder to capture local semantic information. In addition, an object-level representation is included to model the regional interaction map. The extracted hierarchical features are then fed to the contextual attention module to adaptively recalibrate the representation space using the local information. Then, they emphasize the informative regions while taking into account the long-range contextual dependency derived by the Transformer module. We validate our method on several large-scale public medical image segmentation datasets and achieve state-of-the-art performance. We have provided the implementation code in https://github.com/rezazad68/TMUnet.

Long Chen, Yuli Wu, Johannes Stegmaier et al.

Designing metrics for evaluating instance segmentation revolves around comprehensively considering object detection and segmentation accuracy. However, other important properties, such as sensitivity, continuity, and equality, are overlooked in the current study. In this paper, we reveal that most existing metrics have a limited resolution of segmentation quality. They are only conditionally sensitive to the change of masks or false predictions. For certain metrics, the score can change drastically in a narrow range which could provide a misleading indication of the quality gap between results. Therefore, we propose a new metric called sortedAP, which strictly decreases with both object- and pixel-level imperfections and has an uninterrupted penalization scale over the entire domain. We provide the evaluation toolkit and experiment code at https://www.github.com/looooongChen/sortedAP.

Long Chen, Weiwen Zhang, Yuli Wu et al.

To date, most instance segmentation approaches are based on supervised learning that requires a considerable amount of annotated object contours as training ground truth. Here, we propose a framework that searches for the target object based on a shape prior. The shape prior model is learned with a variational autoencoder that requires only a very limited amount of training data: In our experiments, a few dozens of object shape patches from the target dataset, as well as purely synthetic shapes, were sufficient to achieve results en par with supervised methods with full access to training data on two out of three cell segmentation datasets. Our method with a synthetic shape prior was superior to pre-trained supervised models with access to limited domain-specific training data on all three datasets. Since the learning of prior models requires shape patches, whether real or synthetic data, we call this framework semi-supervised learning.

Yuli Wu, Ivan Karetic, Johannes Stegmaier et al.

We propose a neural network-based framework to optimize the perceptions simulated by the in silico retinal implant model pulse2percept. The overall pipeline consists of a trainable encoder, a pre-trained retinal implant model and a pre-trained evaluator. The encoder is a U-Net, which takes the original image and outputs the stimulus. The pre-trained retinal implant model is also a U-Net, which is trained to mimic the biomimetic perceptual model implemented in pulse2percept. The evaluator is a shallow VGG classifier, which is trained with original images. Based on 10,000 test images from the MNIST dataset, we show that the convolutional neural network-based encoder performs significantly better than the trivial downsampling approach, yielding a boost in the weighted F1-Score by 36.17% in the pre-trained classifier with 6x10 electrodes. With this fully neural network-based encoder, the quality of the downstream perceptions can be fine-tuned using gradient descent in an end-to-end fashion.

Long Chen, Yuli Wu, Dorit Merhof

Instance segmentation aims to delineate each individual object of interest in an image. State-of-the-art approaches achieve this goal by either partitioning semantic segmentations or refining coarse representations of detected objects. In this work, we propose a novel approach to solve the problem via object layering, i.e. by distributing crowded, even overlapping objects into different layers. By grouping spatially separated objects in the same layer, instances can be effortlessly isolated by extracting connected components in each layer. In comparison to previous methods, our approach is not affected by complex object shapes or object overlaps. With minimal post-processing, our method yields very competitive results on a diverse line of datasets: C. elegans (BBBC), Overlapping Cervical Cells (OCC) and cultured neuroblastoma cells (CCDB). The source code is publicly available.

Reza Azad, Moein Heidary, Kadir Yilmaz et al.

Semantic image segmentation, the process of classifying each pixel in an image into a particular class, plays an important role in many visual understanding systems. As the predominant criterion for evaluating the performance of statistical models, loss functions are crucial for shaping the development of deep learning-based segmentation algorithms and improving their overall performance. To aid researchers in identifying the optimal loss function for their particular application, this survey provides a comprehensive and unified review of $25$ loss functions utilized in image segmentation. We provide a novel taxonomy and thorough review of how these loss functions are customized and leveraged in image segmentation, with a systematic categorization emphasizing their significant features and applications. Furthermore, to evaluate the efficacy of these methods in real-world scenarios, we propose unbiased evaluations of some distinct and renowned loss functions on established medical and natural image datasets. We conclude this review by identifying current challenges and unveiling future research opportunities. Finally, we have compiled the reviewed studies that have open-source implementations on our GitHub page.

Yuli Wu, Weidong He, Dennis Eschweiler et al.

Modern biomedical image analysis using deep learning often encounters the challenge of limited annotated data. To overcome this issue, deep generative models can be employed to synthesize realistic biomedical images. In this regard, we propose an image synthesis method that utilizes denoising diffusion probabilistic models (DDPMs) to automatically generate retinal optical coherence tomography (OCT) images. By providing rough layer sketches, the trained DDPMs can generate realistic circumpapillary OCT images. We further find that more accurate pseudo labels can be obtained through knowledge adaptation, which greatly benefits the segmentation task. Through this, we observe a consistent improvement in layer segmentation accuracy, which is validated using various neural networks. Furthermore, we have discovered that a layer segmentation model trained solely with synthesized images can achieve comparable results to a model trained exclusively with real images. These findings demonstrate the promising potential of DDPMs in reducing the need for manual annotations of retinal OCT images.

Rüveyda Yilmaz, Kaan Keven, Yuli Wu et al.

Automated cell segmentation in microscopy images is essential for biomedical research, yet conventional methods are labor-intensive and prone to error. While deep learning-based approaches have proven effective, they often require large annotated datasets, which are scarce due to the challenges of manual annotation. To overcome this, we propose a novel framework for synthesizing densely annotated 2D and 3D cell microscopy images using cascaded diffusion models. Our method synthesizes 2D and 3D cell masks from sparse 2D annotations using multi-level diffusion models and NeuS, a 3D surface reconstruction approach. Following that, a pretrained 2D Stable Diffusion model is finetuned to generate realistic cell textures and the final outputs are combined to form cell populations. We show that training a segmentation model with a combination of our synthetic data and real data improves cell segmentation performance by up to 9\% across multiple datasets. Additionally, the FID scores indicate that the synthetic data closely resembles real data. The code for our proposed approach will be available at https://github.com/ruveydayilmaz0/cascaded_diffusion.

Kun Wang, Meng Chen, Junhao Wang et al.

With the widespread deployment of deep-learning-based speech models in security-critical applications, backdoor attacks have emerged as a serious threat: an adversary who poisons a small fraction of training data can implant a hidden trigger that controls the model's output while preserving normal behavior on clean inputs. Existing inference-time defenses are not well suited to the audio domain, as they either rely on trigger over-robustness assumptions that fail on transformation-based and semantic triggers, or depend on properties specific to image or text modalities. In this paper, we propose STEP (Stability-based Trigger Exposure Profiling), a black-box, retraining-free backdoor detector that operates under hard-label-only access. Its core idea is to exploit a characteristic dual anomaly of backdoor triggers: anomalous label stability under semantic-breaking perturbations, and anomalous label fragility under semantic-preserving perturbations. STEP profiles each test sample with two complementary perturbation branches that target these two properties respectively, scores the resulting stability features with one-class anomaly detectors trained on benign references, and fuses the two scores via unsupervised weighting. Extensive experiments across seven backdoor attacks show that STEP achieves an average AUROC of 97.92% and EER of 4.54%, substantially outperforming state-of-the-art baselines, and generalizes across model architectures, speech tasks, an open-set verification scenario, and over-the-air physical-world settings.

Friedhelm Hamann, Emil Mededovic, Fabian Gülhan et al.

We present an overview of the Spatio-temporal Instance Segmentation (SIS) challenge held in conjunction with the CVPR 2025 Event-based Vision Workshop. The task is to predict accurate pixel-level segmentation masks of defined object classes from spatio-temporally aligned event camera and grayscale camera data. We provide an overview of the task, dataset, challenge details and results. Furthermore, we describe the methods used by the top-5 ranking teams in the challenge. More resources and code of the participants' methods are available here: https://github.com/tub-rip/MouseSIS/blob/main/docs/challenge_results.md

Rüveyda Yilmaz, Yuli Wu, Johannes Stegmaier et al.

Positron Emission Tomography (PET) image reconstruction is inherently challenged by Poisson noise and physical degradation factors, which are further exacerbated in limited-angle acquisitions. While deep learning methods demonstrate promising performance, their generalization to unseen clinical data distributions remains limited without extensive retraining. We propose PET-Adapter, a test-time domain adaptation framework for generative PET reconstruction models pretrained solely on phantom data. Our method enables adaptation to clinical datasets with varying anatomies, tracers, and scanner configurations without requiring paired ground truth. PET-Adapter introduces layer-wise low-rank anatomical conditioning during adaptation and Ordered Subset Expectation Maximization-based warm-starting that initializes the generation from physics-informed reconstructions, reducing diffusion steps from 50 to 2 without compromising quality. Experiments across multiple clinical datasets demonstrate superior 3D reconstruction performance in both full-angle and limited-angle settings, highlighting the clinical feasibility and computational efficiency of the proposed approach.

Yuli Wu

This note provides a chronological account of Fréchet distances, starting with Maurice Fréchet's 1906 doctoral thesis on distances in abstract sets and tracing the Fréchet distance between polygonal curves and its algorithmic computation in the 1990s. It then continues with his 1957 paper on a coupling-based distance between probability laws with a brief glimpse of Wasserstein distance and optimal transport. We further attempt to draw connections between the distributional, coupling-based facet of Fréchet distances on probability laws and the geometric facet on curves. The note ends with a modern use case, the Fréchet Inception Distance (FID) in the era of deep generative model evaluation, interpretable as the Wasserstein-2 distance between multivariate Gaussians in a learned feature space. An appendix includes \TeX{}ified faithful English translations of Fréchet's 1906 thesis and 1957 paper, and Lévy's 1950 note for reader convenience.

Yuli Wu, Fucheng Liu, Rüveyda Yilmaz et al.

Fréchet Inception Distance (FID), computed with an ImageNet pretrained Inception-v3 network, is widely used as a state-of-the-art evaluation metric for generative models. It assumes that feature vectors from Inception-v3 follow a multivariate Gaussian distribution and calculates the 2-Wasserstein distance based on their means and covariances. While FID effectively measures how closely synthetic data match real data in many image synthesis tasks, the primary goal in biomedical generative models is often to enrich training datasets ideally with corresponding annotations. For this purpose, the gold standard for evaluating generative models is to incorporate synthetic data into downstream task training, such as classification and segmentation, to pragmatically assess its performance. In this paper, we examine cases from retinal imaging modalities, including color fundus photography and optical coherence tomography, where FID and its related metrics misalign with task-specific evaluation goals in classification and segmentation. We highlight the limitations of using various metrics, represented by FID and its variants, as evaluation criteria for these applications and address their potential caveats in broader biomedical imaging modalities and downstream tasks.

Emil Mededovic, Valdy Laurentius, Yuli Wu et al.

We analyze the capabilities of foundation models addressing the tedious task of generating annotations for animal tracking. Annotating a large amount of data is vital and can be a make-or-break factor for the robustness of a tracking model. Robustness is particularly crucial in animal tracking, as accurate tracking over long time horizons is essential for capturing the behavior of animals. However, generating additional annotations using foundation models can be counterproductive, as the quality of the annotations is just as important. Poorly annotated data can introduce noise and inaccuracies, ultimately compromising the performance and accuracy of the trained model. Over-reliance on automated annotations without ensuring precision can lead to diminished results, making careful oversight and quality control essential in the annotation process. Ultimately, we demonstrate that a thoughtful combination of automated annotations and manually annotated data is a valuable strategy, yielding an IDF1 score of 80.8 against blind usage of SAM2 video with an IDF1 score of 65.6.

Yuli Wu, Julian Wittmann, Peter Walter et al.

Implantable retinal prostheses offer a promising solution to restore partial vision by circumventing damaged photoreceptor cells in the retina and directly stimulating the remaining functional retinal cells. However, the information transmission between the camera and retinal cells is often limited by the low resolution of the electrode array and the lack of specificity for different ganglion cell types, resulting in suboptimal stimulations. In this work, we propose to utilize normalizing flow-based conditional invertible neural networks to optimize retinal implant stimulation in an unsupervised manner. The invertibility of these networks allows us to use them as a surrogate for the computational model of the visual system, while also encoding input camera signals into optimized electrical stimuli on the electrode array. Compared to other methods, such as trivial downsampling, linear models, and feed-forward convolutional neural networks, the flow-based invertible neural network and its conditional extension yield better visual reconstruction qualities w.r.t. various metrics using a physiologically validated simulation tool.

Fabian Gülhan, Emil Mededovic, Yuli Wu et al.

Despite progress toward end-to-end tracking with transformer architectures, poor detection performance and the conflict between detection and association in a joint architecture remain critical concerns. Recent approaches aim to mitigate these issues by (i) employing advanced denoising or label assignment strategies, or (ii) incorporating detection priors from external object detectors via distillation or anchor proposal techniques. Inspired by the success of integrating detection priors and by the key insight that MOTR-like models are secretly strong detection models, we introduce SelfMOTR, a novel tracking transformer that relies on self-generated detection priors. Through extensive analysis and ablation studies, we uncover and demonstrate the hidden detection capabilities of MOTR-like models, and present a practical set of tools for leveraging them effectively. On DanceTrack, SelfMOTR achieves strong performance, competing with recent state-of-the-art end-to-end tracking methods.

Emil Mededovic, Yuli Wu, Henning Konermann et al.

Analyzing animal behavior from video recordings is crucial for scientific research, yet manual annotation remains labor-intensive and prone to subjectivity. Efficient segmentation methods are needed to automate this process while maintaining high accuracy. In this work, we propose a novel pipeline that utilizes eye-tracking data from Aria glasses to generate prompt points, which are then used to produce segmentation masks via a fast zero-shot segmentation model. Additionally, we apply post-processing to refine the prompts, leading to improved segmentation quality. Through our approach, we demonstrate that combining eye-tracking-based annotation with smart prompt refinement can enhance segmentation accuracy, achieving an improvement of 70.6% from 38.8 to 66.2 in the Jaccard Index for segmentation results in the rats dataset.

Rüveyda Yilmaz, Zhu Chen, Yuli Wu et al.

Cell microscopy data are abundant; however, corresponding segmentation annotations remain scarce. Moreover, variations in cell types, imaging devices, and staining techniques introduce significant domain gaps between datasets. As a result, even large, pretrained segmentation models trained on diverse datasets (source datasets) struggle to generalize to unseen datasets (target datasets). To overcome this generalization problem, we propose CellStyle, which improves the segmentation quality of such models without requiring labels for the target dataset, thereby enabling zero-shot adaptation. CellStyle transfers the attributes of an unannotated target dataset, such as texture, color, and noise, to the annotated source dataset. This transfer is performed while preserving the cell shapes of the source images, ensuring that the existing source annotations can still be used while maintaining the visual characteristics of the target dataset. The styled synthetic images with the existing annotations enable the finetuning of a generalist segmentation model for application to the unannotated target data. We demonstrate that CellStyle significantly improves zero-shot cell segmentation performance across diverse datasets by finetuning multiple segmentation models on the style-transferred data. The code will be made publicly available.

Yuli Wu, Do Dinh Tan Nguyen, Henning Konermann et al.

This study proposes a retinal prosthetic simulation framework driven by visual fixations, inspired by the saccade mechanism, and assesses performance improvements through end-to-end optimization in a classification task. Salient patches are predicted from input images using the self-attention map of a vision transformer to mimic visual fixations. These patches are then encoded by a trainable U-Net and simulated using the pulse2percept framework to predict visual percepts. By incorporating a learnable encoder, we aim to optimize the visual information transmitted to the retinal implant, addressing both the limited resolution of the electrode array and the distortion between the input stimuli and resulting phosphenes. The predicted percepts are evaluated using the self-supervised DINOv2 foundation model, with an optional learnable linear layer for classification accuracy. On a subset of the ImageNet validation set, the fixation-based framework achieves a classification accuracy of 87.72%, using computational parameters based on a real subject's physiological data, significantly outperforming the downsampling-based accuracy of 40.59% and approaching the healthy upper bound of 92.76%. Our approach shows promising potential for producing more semantically understandable percepts with the limited resolution available in retinal prosthetics.

Yuli Wu, Peter Walter, Dorit Merhof

Fovea localization is one of the most popular tasks in ophthalmic medical image analysis, where the coordinates of the center point of the macula lutea, i.e. fovea centralis, should be calculated based on color fundus images. In this work, we treat the localization problem as a classification task, where the coordinates of the x- and y-axis are considered as the target classes. Moreover, the combination of the softmax activation function and the cross entropy loss function is modified to its multiscale variation to encourage the predicted coordinates to be located closely to the ground-truths. Based on color fundus photography images, we empirically show that the proposed multiscale softmax cross entropy yields better performance than the vanilla version and than the mean squared error loss with sigmoid activation, which provides a novel approach for coordinate regression.

Yuli Wu, Yucheng Hu, Suting Miao

We present an object detection based approach to localize handwritten regions from documents, which initially aims to enhance the anonymization during the data transmission. The concatenated fusion of original and preprocessed images containing both printed texts and handwritten notes or signatures are fed into the convolutional neural network, where the bounding boxes are learned to detect the handwriting. Afterwards, the handwritten regions can be processed (e.g. replaced with redacted signatures) to conceal the personally identifiable information (PII). This processing pipeline based on the deep learning network Cascade R-CNN works at 10 fps on a GPU during the inference, which ensures the enhanced anonymization with minimal computational overheads. Furthermore, the impressive generalizability has been empirically showcased: the trained model based on the English-dominant dataset works well on the fictitious unseen invoices, even in Chinese. The proposed approach is also expected to facilitate other tasks such as handwriting recognition and signature verification.

Yuli Wu, Long Chen, Dorit Merhof

As a proposal-free approach, instance segmentation through pixel embedding learning and clustering is gaining more emphasis. Compared with bounding box refinement approaches, such as Mask R-CNN, it has potential advantages in handling complex shapes and dense objects. In this work, we propose a simple, yet highly effective, architecture for object-aware embedding learning. A distance regression module is incorporated into our architecture to generate seeds for fast clustering. At the same time, we show that the features learned by the distance regression module are able to promote the accuracy of learned object-aware embeddings significantly. By simply concatenating features of the distance regression module to the images as inputs of the embedding module, the mSBD scores on the CVPPP Leaf Segmentation Challenge can be further improved by more than 8% compared to the identical set-up without concatenation, yielding the best overall result amongst the leaderboard at CodaLab.