Automated melanoma detection in total body photography (TBP) datasets presents a dual challenge: extreme class imbalance (0.1 γ ^+=4.0 , γ ^-=1.0 ), the proposed model achieves a partial AUC (pAUC, FPR ≤ 0.20 ) of 0.1892 on the ISIC 2024 SLICE-3D validation set, exceeding the competition winner (EVA02+GBDT, pAUC =0.1755 ) by 7.8 r=0.6424 , p < 10^-10 ), though the gate’s absolute reliance on ViT features remains small ( <0.03 ) for both classes, with CNN features dominating fusion overall. Ablation demonstrates a 9.1 χ ^2=953.30 , p=2.54× 10^-209 vs. A1; χ ^2=2175.07 , p≈ 0 vs. A3), though this difference is not significant when restricted to the 77 malignant validation lesions alone ( p>0.05 ); a Wilcoxon signed-rank test confirms significantly higher malignant scores overall ( W=40,015,302 , p≈ 0 ).
Cardiovascular disease remains a major contributor to global morbidity, while clinical data sharing across institutions is constrained by privacy regulations such as the GDPR. This retrospective proof-of-concept study investigates whether federated learning (FL) can support collaborative cardiovascular risk prediction without transferring raw patient records across institutional boundaries. Using 46,520 records from the MIMIC-III database, partitioned at the patient level, a centralised neural network is compared with Federated Averaging (FedAvg) and Federated Proximal (FedProx) across three simulated hospital clients under IID and Non-IID distributions. The findings show that federated models achieve predictive performance comparable to centralised training and remain robust under heterogeneous data distributions. Calibration analysis further indicates that distributed aggregation preserves reliable risk estimates for clinical screening. Validation-based operating-point optimisation demonstrates that the models can prioritise high screening sensitivity without additional communication overhead. While local data retention supports data minimisation and reduces cross-institutional data transfer, it does not provide formal privacy guarantees such as differential privacy. Overall, the findings demonstrate the feasibility of FL for privacy-conscious cardiovascular risk prediction and provide a foundation for future evaluation in real-world multi-institutional clinical settings.
Sarcasm detection is difficult because intended meaning often contradicts literal wording, which misleads sentiment analysis and other NLP systems. Although transformer encoders such as RoBERTa capture context well, they can miss subtle contrastive cues. This paper proposes an incongruity-aware framework, IDL-RoBERTa, that injects an interpretable sentiment-contrast signal into a RoBERTa-based classifier. Using VADER, we compute sentence-level polarity and token-level polarity, derive an incongruity feature from their disagreement, and align it with the token sequence as an additional feature channel. IDL-RoBERTa is trained with a multi-task objective that combines sarcasm classification with auxiliary regression of sentence-level incongruity to encourage attention to contrast rather than surface lexical patterns. Model parameters are optimized using gradient-based learning (mini-batch gradient descent) with an AdamW optimizer and a warmup schedule, as configured in the final notebooks. Experiments on the News Headlines Sarcasm Dataset v2 and a Twitter sarcasm dataset show consistent improvements over classical CNN/LSTM baselines and a strong RoBERTa baseline in our recorded runs: headline test F1 rises from 0.8019 to 0.8966, and Twitter test F1 from 0.6541 to 0.7342. We also report back-translation ablations on Twitter to examine augmentation effects. Overall, the approach improves performance while preserving interpretability through an explicit, inspectable proxy for incongruity.
Retinal fundus ultrasound-derived CIMT is well-established as a reliable surrogate measure of subclinical atherosclerosis and cardiovascular risk, but its acquisition demands specialized equipment and expertise, limiting its scalability for mass screening. Retinal fundus image analysis and demography can be considered promising complementary methods to predict carotid IMT. In this study, RetinaFusionNet is presented, a novel multimodal deep learning framework combining bilateral fundus images and demography through a learned gating system that adaptively modifies demographic data prior to fusion. All experiments were performed using the publicly available China-Fundus-CIMT dataset containing 2,903 patients according to its officially released patient-level train, validation, and test split. EfficientNet-B0, ResNet50, and DenseNet 121 were investigated in a common-weights dual-eye architecture including automated retina cropping, advanced data augmentation, label smoothing, class-weighted loss, horizontal flip test-time augmentation, optimized thresholds on validation, and early stopping. The best gated model, EfficientNet-B0, reached an accuracy of 89.00
