Recent advances in Multimodal Large Language Models (MLLMs) have spurred significant progress in Chain-of-Thought (CoT) reasoning. Building on the success of Deepseek-R1, researchers extended multimodal reasoning to post-training paradigms based on reinforcement learning (RL), focusing predominantly on mathematical datasets. However, existing post-training paradigms tend to neglect two critical aspects: (1) The lack of quantifiable difficulty metrics capable of strategically screening samples for post-training optimization. (2) Suboptimal post-training paradigms that fail to jointly optimize perception and reasoning capabilities. To address this gap, we propose two novel difficulty-aware sampling strategies: Progressive Image Semantic Masking (PISM) quantifies sample hardness through systematic image degradation, while Cross-Modality Attention Balance (CMAB) assesses cross-modal interaction complexity via attention distribution analysis. Leveraging these metrics, we design a hierarchical training framework that incorporates both GRPO-only and SFT+GRPO hybrid training paradigms, and evaluate them across six benchmark datasets. Experiments demonstrate consistent superiority of GRPO applied to difficulty-stratified samples compared to conventional SFT+GRPO pipelines, indicating that strategic data sampling can obviate the need for supervised fine-tuning while improving model accuracy.

Prototype-based personalized federated learning methods have emerged as a promising strategy due to their ability to represent client-specific class characteristics effectively through learned class prototypes. These prototypes capture salient features of client-local data, facilitating personalized model adaptation. However, existing prototype-based aggregation strategies predominantly rely on weighted averaging, implicitly assuming prototype consistency across clients. This assumption neglects the intrinsic heterogeneity and non-independent and identically distributed (non-IID) nature of client data, compelling diverse local prototypes to align toward a singular global prototype and consequently causing significant aggregation bias. Motivated by observations from intra-class feature saliency analysis, we identify that clients inherently emphasize distinct feature regions even for the same class. To leverage this intra-class diversity, we introduce FedIC, a novel prototype clustering and collaborative classifier optimization approach. Specifically, FedIC first clusters prototypes based on intra-class similarity to form intra-class prototype subspaces, ensuring that aggregation occurs exclusively within each cluster, thus eliminating the bias stemming from forced global unification. To further exploit the benefits of intra-cluster collaboration, we quantify the combined predictive gains of classifiers from clients within the same cluster as a function of classifier combination weights. This targeted aggregation and collaborative optimization strategy effectively circumvents the bias introduced by global alignment. Extensive experiments under various non-IID settings show that FedIC significantly outperforms existing Prototype-based and Clustered PFL Methods.

The limited availability of high-quality training data poses a persistent challenge for synthetic aperture radar (SAR) target classification. Existing data augmentation methods mainly adopt a simplistic application of GAN-based style transfer techniques to directly synthesize pseudo-SAR images from optical images. However, our in-depth analysis of this cross-modal conversion reveals that such straightforward strategies primarily focus on transferring high-level semantic information (e.g., target shapes), thus failing to adequately capture the essential low-level features unique to SAR imagery (e.g., scattering textures). To address this inherent trade-off between high-level semantic preservation and low-level feature authenticity, we propose a Hierarchical Feature-Constrained GAN (HiFC-GAN) tailored for optical-to-SAR style transfer. Specifically, HiFC-GAN enhances the representation of low-level SAR features by introducing local texture contrast constraints at shallow layers, while introducing explicit feature mapping constraints at deeper layers to maintain high-level semantic consistency throughout the reconstruction process. Experimental results demonstrate that HiFC-GAN significantly outperforms existing GAN-based techniques in image generation quality, particularly improving the low-level feature authenticity of pseudo-SAR images. Moreover, the generated pseudo-SAR images further improve the performance of downstream target classification tasks, yielding accuracy gains ranging from 3.56% to 5.90% on average with mainstream CNN-based models.

Achieving fairness in clinical AI remains challenging due to class bias in long-tailed datasets. We propose Class wise Fair Test-time Adaptation (CFTA), a pragmatic setting where biased pre-trained models are adapted to single test instances without source data. To this end, we introduce FairBN, which hierarchically recalibrates BatchNorm statistics through three modules: Instance-Specific Adaptation (ISA), Memory Bank Maintenance (MBM), and Class-Aware Calibration (CAC). Experiments on four medical benchmarks show that FairBN consistently outperforms recent TTA methods (TENT, CoTTA, RoTTA), achieving notable improvements in both accuracy and class-wise F1 fairness.

Vascular segmentation in medical images is crucial for disease diagnosis and surgical navigation. However, the segmented vascular structure is often discontinuous due to its slender nature and inadequate prior modeling. In this paper, we propose a novel Serpentine Window Mamba (SWinMamba) to achieve accurate vascular segmentation. The proposed SWinMamba innovatively models the continuity of slender vascular structures by incorporating serpentine window sequences into bidirectional state space models. The serpentine window sequences enable efficient feature capturing by adaptively guiding global visual context modeling to the vascular structure. Specifically, the Serpentine Window Tokenizer (SWToken) adaptively splits the input image using overlapping serpentine window sequences, enabling flexible receptive fields (RFs) for vascular structure modeling. The Bidirectional Aggregation Module (BAM) integrates coherent local features in the RFs for vascular continuity representation. In addition, dual-domain learning with Spatial-Frequency Fusion Unit (SFFU) is designed to enhance the feature representation of vascular structure. Extensive experiments on three challenging datasets demonstrate that the proposed SWinMamba achieves superior performance with complete and connected vessels.

Class Incremental Learning (CIL) aims to enable models to continuously learn new categories while retaining previous classification abilities. In medical scenarios, where new disease categories frequently emerge, CIL becomes crucial. Traditional CIL approaches often face "catastrophic forgetting". Analytical Class Incremental Learning (ACIL) offers an analytical (i.e., closed-form) linear solution that does not depend on conventional replay or regularization techniques, thereby mitigating forgetting and addressing privacy concerns, making it suitable for medical datasets. However, few studies have explored the problem of knowledge forgetting in CIL for medical data with ACIL. Based on the latest research, we systematically study this problem for the first time. Specifically, we present a benchmark named AIDC (Analytical Incremental Disease Classification), comparing ACIL against five established CIL methods across three medical datasets. Results show that ACIL achieves notably higher average classification accuracy and exhibits better anti-forgetting capabilities compared to traditional methods.

Semi-supervised medical image segmentation faces two challenging issues: (1) insufficient exploration of latent structures leading to difficulty in comprehensively capturing complex features and structures in medical images; (2) sensitivity to noise, where unlabeled data lacks accurate label information, making the model more prone to noise interference during the learning process. In this paper, a method, uncertainty-aware consistency learning (UAC), is proposed to improve the poor generalization and suboptimal performance in semi-supervised medical image segmentation caused by insufficient information exploration and sensitivity to noise. Firstly, by employing multiple perturbation strategies at both the input and output levels, specifically through data-level and scale-level perturbations, the model is better equipped to capture structural information within organs and essential features that impact segmentation performance . Secondly, the perturbation uncertainty leverages perturbation prediction differences to measure uncertainty helps the model generate reliable predictions and avoid excessive focus on unreliable areas in the predictions. Experimental results on three public medical image segmentation datasets demonstrate that our UAC, utilizing multiple perturbation strategies and uncertainty estimation , exhibits generality across various organ segmentation tasks and achieves accurate segmentation, with the DICE of 91.15%(LA), 77.52%(Pancreas-CT) and 78.71%(PARSE) under a 10% label ratio setting. Comparative and ablation studies indicate that our method outperforms state-of-the-art semi-supervised medical image segmentation methods .

Abnormal retinal vascular morphology is commonly associated with cardiac, cerebrovascular, and systemic diseases. Hence, automated artery/vein(A/V) classification is crucial for the diagnosis of ophthalmic and systemic diseases. However, existing methods still face limitations in A/V classification and are prone to errors especially in microvessels and in noisy backgrounds. To alleviate these problems, this paper proposes an Evidence Probability Network (EP-Net) to achieve accurate A/V classification. Concretely, the multi-scale feature module in the EP-Net learns various vessel features, and the evidence probability module measures uncertainty and evidence for each pixel to overcome misclassification because of over-/under-confidence. Experiments on two public fundus image datasets demonstrate the superiority of the proposed EP-Net over state-of-the-art A/V classification methods.

Computed tomography (CT) scans are clinically important in radiotherapy planning (RTP) for ROI contour delineation and radiation dose calculation. However, it involves significant radiation exposure, which can bring potential health problems. Nowadays, the synthesis of MR to CT provides an alternative to repetitive CT examinations. Although transformers are widely used for image synthesis, achieving effective multi-regions pseudo-CT synthesis from magnetic resonance (MR) images faces common and unique challenges: 1) The quadratic time complexity problem. While transformer facilitates long-range modeling in image synthesis, efficiently integrating the self-attention mechanism with 3D volume data remains an unresolved challenge. 2) The modal differences between MR and CT are significant, and the complex structural priors present within MR and CT make it difficult for transformer to learn effective mapping functions. To address these issues, we propose a high frequency-information guided net for multi-regions pseudo-CT synthesis (HFGS), efficiently generating multi-regions pseudo-CTs from different MR sequences. Our carefully designed 3D cascaded frequency transformer (CFT) serves as the synthesis module, utilizing element-wise product of frequency domain signals instead of matrix multiplication in spatial domain. This approach provides an efficient self-attention calculation method and improves synthesis efficiency. Additionally, to tackle the challenge of insufficient high-frequency information for high-quality decoding, we have designed a learnable guidance module to capture important structural priors within both the source and target modalities, guiding the synthesis module to produce high-quality pseudo-CTs. Our code will be available at https://github.com/qijianyu277/HFGS.

Discovering visual evidence is of great importance in making glaucoma diagnosis trustworthy, with interpretable processes and reliable results. Existing works usually learn image representation for glaucoma diagnosis, where the cup-to-disc ratios (CDRs) are employed as the quantitative evidence to interpret the diagnosis model. However, they rely on global visual features which are insufficient for the trustworthy interpretation of medical image for disease diagnosis. To enable interpretable and reliable glaucoma diagnosis, this paper proposes confidence-guided weakly-supervised learning (CG-WSL) by exploiting the intrinsic visual-semantic interactions in fundus images. The CG-WSL discovers the evidential local regions to support the reliable glaucoma diagnosis with fine-grained anatomical representations, only given the image-level annotations. The evidential local regions not only provide the localization information about the lesions and anatomies for visual interpretation, but also enhance the feature presentation with fine-grained anatomy-level features to discriminate the abnormal cases. Specifically, it consists of two parts: confidence-guided evidence discovery (CGED) for multi-scale fine-grained visual evidence discovery, and feature weighted fusion (FWF) for coarse-to-fine grained representation learning. Experimental results on two datasets demonstrate the effectiveness of proposed method on glaucoma diagnosis with accuracy of 0.981 (LAG) and 0.956 (RIM-ONE r2). Visualization results indicate the visual evidence for glaucoma diagnosis, which makes the diagnosis process interpretable and results reliable.

Existing computer-aided diagnosis models only focus on the statistical accuracy for glaucoma diagnosis, but ignore its reliability for the predictions. Reliability refers to the degree how the model’s prediction can be trusted when be used to assess glaucoma. Predictions with higher reliability can help the clinician make a confident decision, while the prediction with lower reliability will confuse the decision-making. Researches indicate that reliability is high related to the uncertainty both from model and data. In this paper, a method, multi-sample evidential deep learning (MS-EBDL), is proposed to enable the glaucoma assessment model with high reliability. The proposed MS-EBDL gives the disease prediction with quantitative confidence to indicate its reliability by capturing sufficient epistemic. Therefore, the proposed MS-EBDL consists of two parts: evidential deep learning for fundamental epistemic uncertainty and multi-sample dropout for additional epistemic uncertainty. Experimental results on two glaucoma datasets shown that the proposed MS-EBDL outperforms the benchmark with 97.75%(LAG) and 92.20%(RIM-ONE-R2) of accuracy, and provides reliable prediction confidence, which helps clinician make the right decision.

Radiation therapy is the primary treatment for recurrent nasopharyngeal carcinoma. However, it may induce necrosis of the nasopharynx, leading to severe complications such as bleeding and headache. Therefore, forecasting necrosis of the nasopharynx and initiating timely clinical intervention has important implications for reducing complications caused by re-irradiation. This research informs clinical decision-making by making predictions on re-irradiation of recurrent nasopharyngeal carcinoma using deep learning multi-modal information fusion between multi-sequence nuclear magnetic resonance imaging and plan dose. Specifically, we assume that the hidden variables of model data can be divided into two categories: task-consistency and task-inconsistency. The task-consistency variables are characteristic variables contributing to target tasks, while the task-inconsistency variables are not apparently helpful. These modal characteristics are adaptively fused when the relevant tasks are expressed through the construction of supervised classification loss and self-supervised reconstruction loss. The cooperation of supervised classification loss and self-supervised reconstruction loss simultaneously reserves the information of characteristic space and controls potential interference simultaneously. Finally, multi-modal fusion effectively fuses information through an adaptive linking module. We evaluated this method on a multi-center dataset. and found the prediction based on multi-modal features fusion outperformed predictions based on single-modal, partial modal fusion or traditional machine learning methods.

Accurate segmentation of the optic disc and cup (OD/OC) in fundus images is crucial for the glaucoma diagnosis and treatment. However, the distribution discrepancies (domain shift) between source domain and target domain hinder the generalization of segmentation models in clinical applications. In this paper, a dual gradient alignment framework (DGDA) for unsupervised domain adaptation is proposed to achieve accurate OD/OC segmentation. Specifically, feature gradient alignment module is well-designed to learn the domain-invariant representation by aligning the feature gradients between source and target domains. Furthermore, task gradient alignment module introduces the meta-learning to learn the task-agnostic representation to simultaneously balance domain adaptation and OD/OC segmentation. Here, meta-learning imitates multi-task training by aligning the gradients between the segmentation and domain classification task. Extensive experiments are conducted on the two public fundus image datasets (Drishti-GS and RIM-ONE-r3) to evaluate the effectiveness of the proposed DGDA. Experimental results demonstrate the DGDA successfully achieves the unsupervised domain adaptive OD/OC segmentation with competitive performance compared with the state-of-the-art.

Color fundus (CF) image registration is crucial for accurate information fusion; it could obtain more details of retinal structure to assist clinical diagnosis. Existing methods suf-fer from costing time or dataset size, making CF image reg-istration still a challenging task. In this paper, we propose a novel pair-metamorphosis-decouple synthetic data scheme for learning-based CF image registration and ameliorate the registration model for retinal image. Specifically, we take ad-vantage of the pairing information of the registration task to decouple the differences between the pairing data and expand the representative ability of the dataset by synthesizing data. Furthermore, the registration framework is ameliorated ac-cording to the characteristics of the blood vessels in the retinal image. Experiments on the public dataset (FIRE) show that our synthetic data scheme could bring general performance promotion to registration models, and our registration method is superior to other state-of-the-art unsupervised algorithms.

Despite the potential to revolutionise disease diagnosis by performing data-driven classification, clinical interpretability of ConvNet remains challenging. In this paper, a novel clinical interpretable ConvNet architecture is proposed not only for accurate glaucoma diagnosis but also for the more transparent interpretation by highlighting the distinct regions recognised by the network. To the best of our knowledge, this is the first work of providing the interpretable diagnosis of glaucoma with the popular deep learning model. We propose a novel scheme for aggregating features from different scales to promote the performance of glaucoma diagnosis, which we refer to as M-LAP. Moreover, by modelling the correspondence from binary diagnosis information to the spatial pixels, the proposed scheme generates glaucoma activations, which bridge the gap between global semantical diagnosis and precise location. In contrast to previous works, it can discover the distinguish local regions in fundus images as evidence for clinical interpretable glaucoma diagnosis. Experimental results, performed on the challenging ORIGA datasets, show that our method on glaucoma diagnosis outperforms state-of-the-art methods with the highest AUC (0.88). Remarkably, the extensive results, optic disc segmentation (dice of 0.9) and local disease focus localization based on the evidence map, demonstrate the effectiveness of our methods on clinical interpretability.

Glaucoma is a chronic eye disease that leads to irreversible vision loss. The Cup-to-Disc Ratio (CDR) serves as the most important indicator for glaucoma screening and plays a significant role in clinical screening and early diagnosis of glaucoma. In general, obtaining CDR is subjected to measuring on manually or automatically segmented optic disc and cup. Despite great efforts have been devoted, obtaining CDR values automatically with high accuracy and robustness is still a great challenge due to the heavy overlap between optic cup and neuroretinal rim regions. In this paper, a direct CDR estimation method is proposed based on the well-designed semi-supervised learning scheme, in which CDR estimation is formulated as a general regression problem while optic disc/cup segmentation is cancelled. The method directly regresses CDR value based on the feature representation of optic nerve head via deep learning technique while bypassing intermediate segmentation. The scheme is a two-stage cascaded approach comprised of two phases: unsupervised feature representation of fundus image with a convolutional neural networks (MFPPNet) and CDR value regression by random forest regressor. The proposed scheme is validated on the challenging glaucoma dataset Direct-CSU and public ORIGA, and the experimental results demonstrate that our method can achieve a lower average CDR error of 0.0563 and a higher correlation of around 0.726 with measurement before manual segmentation of optic disc/cup by human experts. Our estimated CDR values are also tested for glaucoma screening, which achieves the areas under curve of 0.905 on dataset of 421 fundus images. The experiments show that the proposed method is capable of state-of-the-art CDR estimation and satisfactory glaucoma screening with calculated CDR value.

Evidence identification, optic disc segmentation and automated glaucoma diagnosis are the most clinically significant tasks for clinicians to assess fundus images. However, delivering the three tasks simultaneously is extremely challenging due to the high variability of fundus structure and lack of datasets with complete annotations. In this paper, we propose an innovative Weakly-Supervised Multi-Task Learning method (WSMTL) for accurate evidence identification, optic disc segmentation and automated glaucoma diagnosis. The WSMTL method only uses weak-label data with binary diagnostic labels (normal/glaucoma) for training, while obtains pixel-level segmentation mask and diagnosis for testing. The WSMTL is constituted by a skip and densely connected CNN to capture multi-scale discriminative representation of fundus structure; a well-designed pyramid integration structure to generate high-resolution evidence map for evidence identification, in which the pixels with higher value represent higher confidence to highlight the abnormalities; a constrained clustering branch for optic disc segmentation; and a fully-connected discriminator for automated glaucoma diagnosis. Experimental results show that our proposed WSMTL effectively and simultaneously delivers evidence identification, optic disc segmentation (89.6% TP Dice), and accurate glaucoma diagnosis (92.4% AUC). This endows our WSMTL a great potential for the effective clinical assessment of glaucoma.

BACKGROUND: Glaucoma is an irreversible eye disease caused by the optic nerve injury. Therefore, it usually changes the structure of the optic nerve head (ONH). Clinically, ONH assessment based on fundus image is one of the most useful way for glaucoma detection. However, the effective representation for ONH assessment is a challenging task because its structural changes result in the complex and mixed visual patterns. METHOD: We proposed a novel feature representation based on Radon and Wavelet transform to capture these visual patterns. Firstly, Radon transform (RT) is used to map the fundus image into Radon domain, in which the spatial radial variations of ONH are converted to a discrete signal for the description of image structural features. Secondly, the discrete wavelet transform (DWT) is utilized to capture differences and get quantitative representation. Finally, principal component analysis (PCA) and support vector machine (SVM) are used for dimensionality reduction and glaucoma detection. RESULTS: The proposed method achieves the state-of-the-art detection performance on RIMONE-r2 dataset with the accuracy and area under the curve (AUC) at 0.861 and 0.906, respectively. CONCLUSION: In conclusion, we showed that the proposed method has the capacity as an effective tool for large-scale glaucoma screening, and it can provide a reference for the clinical diagnosis on glaucoma.

Automatic optic disc (OD) localisation and segmentation is still a great challenge in computer‐aided diagnosis and screening system. Here, a new OD segmentation algorithm is proposed based on the distinct features of OD in terms of its intensity and shape. The algorithm includes four stages: image preprocessing, image segmentation, ellipse fitting, and OD localisation and segmentation. In the preprocessing stage, the blood vessel in the input retinal image is removed by using the morphological operation and median filtering in HSL (hue–saturation–lightness) colour space. In the image segmentation and ellipse fitting stages, the fractional‐order Darwinian particle swarm optimisation algorithm is used to extract the brightest region, and the least‐squares optimisation is adopted to detect elliptical OD shape. Finally, the smooth OD borders are generated in the last stage. The proposed method is evaluated by the centroid difference, overlapping ratio, overlap score, and success indexes. Experimental results on the retinal images from DRION, MESSIDOR, ORIGA, and many other public databases demonstrate that the proposed method has superior performance, and may be a suitable tool for automated retinal image analysis.

Contour is a critical feature for image description and object recognition in many computer vision tasks. However, detection of object contour remains a challenging problem because of disturbances from texture edges. This letter proposes a scheme to handle texture edges by implementing contour integration. The proposed scheme integrates structural segments into contours while inhibiting texture edges with the help of the orientation histogram-based center-surround interaction model. In the model, local edges within surroundings exert a modulatory effect on central contour cues based on the co-occurrence statistics of local edges described by the divergence of orientation histograms in the local region. We evaluate the proposed scheme on two well-known challenging boundary detection data sets (RuG and BSDS500). The experiments demonstrate that our scheme achieves a high [Formula: see text]-measure of up to 0.74. Results show that our scheme achieves integrating accurate contour while eliminating most of texture edges, a novel approach to long-range feature analysis.