Identifying gene co-expression (GCE) modules from transcriptomic data is critical for understanding cellular functions and disease mechanisms. Traditional methods often rely on rigid assumptions to detect modules with specific topologies (e.g., spherical clusters or fully connected cliques), limiting their ability to capture the complex and heterogeneous co-expression patterns across different cancer types. Graph convolutional networks (GCNs) have been shown to encode topological structures and node content of networks, which provide a promising direction for more flexible gene co-expression modeling. However, directly applying naïve GCNs to detect GCE modules poses several challenges: sensitivity to expression data noise, lack of explicit mechanisms to obtain arbitrary-shaped modules, and lack of biological ground-truth for supervision. In this study, we propose GraphSC, a graph-based semi-supervised framework for flexible GCE module detection across pan-cancers. To reduce noise sensitivity, GraphSC adopts a graph attention-based encoder-decoder GCN for simultaneous gene embedding learning and expression data reconstruction. GraphSC also integrates a graph-theoretic spectral clustering to organize embeddings into flexible modules without shape constraints. Moreover, we design three loss functions to jointly optimize GraphSC: an unsupervised reconstruction loss for denoising, a structure preservation loss to retain local topology, and a phenotype-supervised module eigengene significance loss to ensure biological relevance. Experiments on eight TCGA datasets demonstrate that GraphSC outperforms state-of-the-art methods in both clustering quality and biological significance. Interpretability analyses further show that GraphSC identifies topologically diverse GCE modules enriched with cancer-specific functions, suggesting its potential for biomarker discovery in different cancers.
Modeling the longitudinal progression of brain atrophy in Alzheimer's Disease (AD) is critical for early diagnosis and prognosis, but remains challenging for existing generative models. Current methods either fail to preserve anatomical integrity or treat progression as an entangled process, where changes from normal aging are inseparably fused with AD pathology. This fundamentally limits their scientific utility. To address this, we propose DeMorph, a novel deformation-based framework that explicitly disentangles and accurately predicts longitudinal brain MRI changes. Our transformer-based architecture features two key innovations: (1) a Decoupled Feed-Forward Network (D-FFN) for separate aging and pathology deformation pathways; and (2) a Multi-Scale Anatomical Guidance Cross-Attention (MS-AGCA) module, which explicitly injects anatomical priors to ensure spatially precise and biologically plausible disentanglement. Extensive experiments on four datasets (ADNI, OASIS-3, AIBL, and an inhouse dataset) demonstrate that DeMorph achieves stateof- the-art performance in both image similarity and anatomical accuracy. Crucially, we validate the success of our framework through robust bidirectional counterfactual synthesis, demonstrating the ability to controllably induce and remove a quantitatively plausible, AD-specific pathological effect.
The subjective nature of descriptive endoscopic diagnosis and the propensity of large vision-language models (LVLMs) to produce semantic hallucinations pose significant threats to clinical safety and standardized patient management. While computer-aided detection (CADe) systems are widely adopted for lesion localization, the transition toward standardized descriptive diagnosis (CADx) remains hindered by large model size, high inference latency, and hallucination risks. To address these barriers, we propose EndoCADx, a compact, low-latency generative framework tailored for standardized gastric descriptive diagnosis. We formalize a two-stage optimization pipeline comprising domain-specific supervised fine-tuning (SFT) and error-driven Direct Preference Optimization (DPO). By systematically mining failure-mode subsets from the initial SFT baseline, we perform targeted recalibration of the model's latent policy to improve concordance between generated descriptions and expert reference annotations. EndoCADx provides an on-demand response time of 1.46 s per physician-selected frozen keyframe, compatible with the routine “freeze-and-inspect” workflow. Extensive evaluation on EndoGastro-500, a multi-center independent cohort from six heterogeneous clinical centers, demonstrates robust performance: the DPO-aligned model achieves 91.6% Lesion Recognition Accuracy (ALR) and 84.7% Feature Description Accuracy (AFD), with the latter reflecting improved attribute-level consistency with expert annotations. A comprehensive Multi-Reader, Multi-Case (MRMC) study involving 12 endoscopists further reveals that EndoCADx acts as a standardizing agent; while conventional CADe alone yields only moderate inter-rater reliability (Kappa = 0.64), EndoCADx elevates diagnostic consistency to almost perfect agreement (Kappa = 0.88), particularly benefiting novice clinicians with a 17.20% accuracy gain. By balancing real-time efficiency with expert-aligned safety, this work offers a practical paradigm for the clinical translation of endoscopic VLMs as a reliable supplementary aid to professional judgment.
Objective: To evaluate whether semantic clinical knowledge augmentation through Semantic Clinical Artificial Intelligence (SCAI) improves medical question answering across commercial and open large language models, and to examine how a reasoning distilled model performs on the same benchmark. Materials and Methods: We evaluated Google Gemini, Microsoft Copilot, Meta Llama 3 70B, and DeepSeek-R1-Distill-Llama-70B on official text only United States Medical Licensing Exam (USMLE) sample questions from Step 1 (n = 87), Step 2 CK (n = 103), and Step 3 (n = 123), with and without SCAI. SCAI comprised HD-NLP parsing, graph and knowledge graph embeddings, a trained SCAI LLM semantic knowledge reasoner, and a semantic knowledge base; the current implementation contained 33,023,902 triples across 246 relation labels. Responses were scored against the answer key with secondary human verification. All models produced a response to every text only item, so incorrect responses were counted as confabulations. Results: Across 313 items, accuracy increased from 90.7% to 97.1% for Gemini, 82.4% to 92.0% for Llama 3 70B, and 55.3% to 81.8% for DeepSeek-R1-Distill; Copilot changed from 93.6% to 94.2%. Within model gains were statistically significant for Gemini, Llama 3 70B, and DeepSeek-R1-Distill on all three steps, but not for Copilot. The best single model result was Llama 3 70B plus SCAI on Step 3 (121/123, 98.4%). Gemini plus SCAI was the numerically strongest commercial system. Conclusions: SCAI was associated with materially better accuracy and lower confabulation (False Positive Results) across several heterogeneous LLMs, supporting a semantically grounded pipeline in which a dedicated SCAI LLM generates clinical context for a host LLM while suggesting that strong general reasoning performance does not automatically translate into healthcare readiness.
Cell instance segmentation in high-resolution microscopy images remains challenging when dense annotations are limited across diverse imaging modalities. This paper presents MUST, a Mean-teacher inspired and Uncertainty-aware Semi-supervised Transformer for cell instance segmentation under limited target-domain annotations. MUST uses adaptive supervised warm-up to stabilize the student model before activating unlabeled consistency learning, and adopts a two-stream mini-batch design to jointly optimize labeled and unlabeled samples. Instead of relying on hard pseudo-labels, an EMA teacher provides soft probability maps as consistency targets for strongly perturbed student predictions. To reduce error propagation from uncertain teacher predictions, MUST further applies entropy- and region-aware reliability weighting together with dynamic unsupervised loss scaling. During inference, sliding-window prediction and test-time augmentation are integrated to improve robustness on high-resolution microscopy images. On the NeurIPS22-CellSeg benchmark, following supervised pre-training on external annotated datasets and using only 10% of the labeled CellSeg training images for supervised optimization, MUST achieves an Overall F1 score of 0.8878 on the official tuning set, reaching the third-best level among representative topperforming challenge methods. These results demonstrate that MUST provides an effective solution for multi-modality cell instance segmentation with reduced target-domain annotation requirements.
