Early diagnosis of neurological disorders (NDs) benefits from multiple complementary signals, including anatomical MRI, diffusion imaging, functional activation, and clinical or serum biomarkers. These modalities capture distinct aspects of neurodegeneration and have motivated recent multimodal learning approaches. However, most existing methods combine modalities only at the feature or decision level through late fusion, meaning that modalities are processed independently and never influence one another during feature extraction. As a result, subtle cross-modality patterns that are highly informative in prodromal stages remain underutilized. To address this limitation, we propose NeuroMoE++, a hierarchical end-to-end framework that enables explicit interaction across modalities while preserving their individual structure. After obtaining unimodal and cross-modality predictions, a subject-driven adaptive integration produces patient-specific decisions. This combination allows the model to capture complementary structure, leverage fused evidence when beneficial, and remain interpretable from a clinical standpoint. Experiments on real clinical datasets demonstrate that NeuroMoE++ substantially outperforms unimodal and late-fusion baselines, achieving 84.94% accuracy. These results highlight the value of explicit cross-modality reasoning for practical neurological diagnosis.
Melanoma is the most lethal form of skin cancer, and early detection is critical for improving patient outcomes. Although dermoscopy combined with deep learning has advanced automated skin-lesion analysis, progress is hindered by limited access to large, well-annotated datasets and by severe class imbalance, where melanoma images are substantially underrepresented. To address these challenges, we present the first systematic benchmarking study comparing four GAN architectures—DCGAN, StyleGAN2, and two StyleGAN3 variants (T and R)—for high-resolution () melanoma-specific synthesis. We train and optimize all models on two expert-annotated benchmarks (ISIC 2018 and ISIC 2020) under unified preprocessing and hyperparameter exploration, with particular attention to R1 regularization tuning. Image quality is assessed through a multi-faceted protocol combining distribution-level metrics (FID), sample-level representativeness (FMD), qualitative dermoscopic inspection, downstream classification with a frozen EfficientNet-based melanoma detector, and independent evaluation by two board-certified dermatologists. StyleGAN2 achieves the best balance of quantitative performance and perceptual quality, attaining FID scores of 24.8 (ISIC 2018) and 7.96 (ISIC 2020) at . The frozen classifier recognizes 83% of StyleGAN2-generated images as melanoma, while dermatologists distinguish synthetic from real images at only 66.5% accuracy (chance = 50%), with low inter-rater agreement (). In a controlled augmentation experiment, adding synthetic melanoma images to address class imbalance improved melanoma detection AUC from 0.925 to 0.945 on a held-out real-image test set. These findings demonstrate that StyleGAN2-generated melanoma images preserve diagnostically relevant features and can provide a measurable benefit for mitigating class imbalance in melanoma-focused machine learning pipelines.
Semantic segmentation is often challenged by occlusions, ambiguous boundaries, and poor lighting conditions when relying on a single frame. While processing auxiliary views can provide valuable spatial–temporal information, doing so naively often introduces significant computational overhead. To tackle this problem, we propose ASTRA: Adaptive Spatial–Temporal Region Augmentation—a fast, low-cost add-on module for integrating auxiliary views to enhance single-frame semantic segmentation. ASTRA's Critical Region Localizer identifies low-confidence, high-uncertainty regions by computing the KL divergence of the segmentation probability distributions. ASTRA then adaptively fuses spatial–temporal information from auxiliary views using a Collaborative Spatial-Temporal Enhancer, which employs sparse attention to refine only those critical regions selectively. To further improve efficiency, we apply a shape-regularized Felzenszwalb algorithm to divide the critical regions into small masks and further reduce the computation required for sparse attention updates. Therefore, ASTRA is both efficient and cost-effective. As a result, ASTRA significantly boosts single-frame segmentation performance, on average improving mIoU by 37.2% and accuracy by 7.5%, with only a 16.1% increase in runtime. Furthermore, ASTRA can be seamlessly integrated into most existing 2D segmentation architectures when auxiliary inputs are available, making it well-suited for applications such as autonomous driving, robotics, and scene monitoring.