Chao Xin

Domain generalization (DG) for medical image segmentation is commonly approached by simulating domain shifts through deterministic image transformations or feature perturbations. However, such methods implicitly assume that unseen domains can be approximated by predefined appearance variations. In this work, we challenge this assumption and hypothesize that medical domain shifts are more fundamentally characterized as uncertainty in spectral distributions. While anatomical structures are largely preserved across institutions, scanners, and acquisition protocols, substantial variations arise in image appearance, texture, and artifacts, which are predominantly manifested in the spectral domain. To test this hypothesis, we propose Spectral-Distribution Uncertainty (SDU), a frequency-aware DG framework that explicitly models uncertainty in the Fourier amplitude distributions of intermediate representations. By decomposing Transformer features into amplitude and phase components, SDU disentangles domain-specific spectral characteristics from domain-invariant anatomical semantics. Instead of imposing handcrafted frequency perturbations, SDU estimates spectral-distribution uncertainty during training and injects stochastic variations into the amplitude space, synthesizing diverse yet anatomically consistent representations that emulate potential unseen domains. Consequently, SDU expands the support of spectral feature distributions, enabling the model to learn representations that are robust to uncertainty-induced spectral shifts. Extensive experiments on retinal fundus and prostate MRI benchmarks demonstrate that SDU consistently outperforms state-of-the-art DG methods. Beyond empirical gains, these findings provide evidence that medical domain shifts are more appropriately characterized as uncertainty in spectral feature distributions, offering a principled perspective for generalizable medical image segmentation.

The rising prevalence of skin lesions places a heavy burden on global health resources and necessitates an early and precise diagnosis for successful treatment. The diagnostic potential of recent multi-modal skin lesion detection algorithms is limited because they ignore dynamic interactions and information sharing across modalities at various feature scales. To address this, we propose a deep learning framework, Multi-Modal Skin-Imaging-based Information-Switching Network (MDSIS-Net), for end-to-end skin lesion recognition. MDSIS-Net extracts intra-modality features using transfer learning in a multi-scale fully shared convolutional neural network and introduces an innovative information-switching module. A cross-attention mechanism dynamically calibrates and integrates features across modalities to improve inter-modality associations and feature representation in this module. MDSIS-Net is tested on clinical disfiguring dermatosis data and the public Derm7pt melanoma dataset. A Visually Intelligent System for Image Analysis (VISIA) captures five modalities: spots, red marks, ultraviolet (UV) spots, porphyrins, and brown spots for disfiguring dermatosis. The model performs better than existing approaches with an mAP of 0.967, accuracy of 0.960, precision of 0.935, recall of 0.960, and f1-score of 0.947. Using clinical and dermoscopic pictures from the Derm7pt dataset, MDSIS-Net outperforms current benchmarks for melanoma, with an mAP of 0.877, accuracy of 0.907, precision of 0.911, recall of 0.815, and f1-score of 0.851. The model’s interpretability is proven by Grad-CAM heatmaps correlating with clinical diagnostic focus areas. In conclusion, our deep multi-modal information-switching model enhances skin lesion identification by capturing relationship features and fine-grained details across multi-modal images, improving both accuracy and interpretability. This work advances clinical decision making and lays a foundation for future developments in skin lesion diagnosis and treatment.