Objectives: Early skin cancer diagnosis can save lives; however, traditional methods rely on expert knowledge and can be time-consuming. This calls for automated systems using machine learning and deep learning. However, existing datasets often focus on flat skin surfaces, neglecting more complex cases on organs or with nearby lesions. Methods: This work addresses this gap by proposing a skin cancer diagnosis methodology using a dataset named ASAN that covers diverse skin cancer cases but suffers from noisy features. To overcome the noisy feature problem, a segmentation dataset named SASAN is introduced, focusing on Region of Interest (ROI) extraction-based classification. This allows models to concentrate on critical areas within the images while ignoring learning the noisy features. Results: Various deep learning segmentation models such as UNet, LinkNet, PSPNet, and FPN were trained on the SASAN dataset to perform segmentation-based ROI extraction. Classification was then performed using the dataset with and without ROI extraction. The results demonstrate that ROI extraction significantly improves the performance of these models in classification. This implies that SASAN is effective in evaluating performance metrics for complex skin cancer cases. Conclusions: This study highlights the importance of expanding datasets to include challenging scenarios and developing better segmentation methods to enhance automated skin cancer diagnosis. The SASAN dataset serves as a valuable tool for researchers aiming to improve such systems and ultimately contribute to better diagnostic outcomes.

Background/Objectives: Accurate classification of primary bone tumors from radiographic images is essential for early diagnosis, appropriate treatment planning, and informed clinical decision-making. While deep convolutional neural networks (CNNs) have shown strong performance in medical image analysis, their high computational complexity often limits real-world clinical deployment. This study aims to develop a lightweight yet highly accurate model for multi-class bone tumor classification. Methods: We propose Bone-CNN, a computationally efficient CNN architecture specifically designed for radiograph-based classification of primary bone tumors. The model was evaluated using the publicly available Figshare Radiograph Dataset of Primary Bone Tumors, which includes nine distinct tumor classes ranging from benign to malignant lesions and originates from multiple imaging centres. Performance was assessed through extensive experiments and compared against established baseline models, including DenseNet121, EfficientNet-B0, and MobileNetV2. Results: Bone-CNN achieved a test accuracy of 96.52% and a macro-AUC of 0.9989, outperforming all baseline architectures. Both quantitative and qualitative evaluations, including confusion matrices and ROC curve analyses, demonstrated robust and reliable discrimination between challenging tumor subtypes. Conclusions: The results indicate that Bone-CNN offers an excellent balance between accuracy and computational efficiency. Its strong performance and lightweight design highlight its suitability for clinical deployment, supporting effective and scalable radiograph-based assessment of primary bone tumors.

Background: Deep-learning systems for endoscopic image classification are commonly evaluated with random data splits, which may overestimate performance under acquisition shift; uncertainty-aware selective prediction may improve reliability by allowing a model to abstain on uncertain cases. Methods: We evaluated binary abnormality detection in two endoscopic imaging domains: laryngeal contact-endoscopy narrow-band imaging (CE-NBI; 210 patients, patient-level) and gastrointestinal endoscopy (HyperKvasir; 6746 images, image-level, predominantly white-light). ImageNet-pretrained ResNet-50 deep ensembles were assessed under a resolution-defined acquisition-shift stress test; for the gastrointestinal data, a random stratified split was additionally used as an in-distribution reference. We evaluated discrimination, calibration, decision-curve analysis, and entropy-based selective prediction. Results: In the gastrointestinal dataset, random-split evaluation produced high performance (AUROC 0.994, 95% CI 0.990–0.996; AUPRC 0.991). Under resolution shift on the same data, performance fell to AUROC 0.723 (95% CI 0.710–0.736; AUPRC 0.690; sensitivity 0.417; specificity 0.874); the two intervals do not overlap. Selective prediction improved reliability among retained cases: under resolution shift, accuracy rose from 0.683 at full coverage to 0.852 (95% CI 0.835–0.871) at 25% coverage (balanced accuracy 0.646 → 0.798). Predictive entropy was significantly higher for incorrect than for correct predictions in both regimes (Mann–Whitney p = 7.1 × 10⁻⁷⁷ with rank-biserial |r| = 0.30 under shift). In the laryngeal cohort, no statistically significant differences were detected among four architectures (ROC-AUC 0.844–0.901; all pairwise DeLong p > 0.05). Conclusions: Random-split evaluation substantially overestimated performance relative to a resolution-defined acquisition-shift stress test, and entropy-based selective prediction improved reliability by identifying a high-confidence subset for automated prediction while deferring the remainder to human review. Target-domain recalibration substantially restores calibration under shift (ECE 0.172 → 0.036 with temperature scaling; → 0.017 with isotonic regression) but does not recover discrimination; selective prediction is complementary, mitigating residual confident-wrong predictions. An encoder-transfer experiment showed asymmetric cross-domain utility; features learned on the larger gastrointestinal cohort transferred to the laryngeal cohort (AUROC 0.80 vs. in-domain 0.89), whereas the reverse direction did not transfer (0.53 vs. 0.72). Prospective multi-center validation remains required before clinical deployment. All code, fold definitions, random seeds, and a reproducible protocol are publicly released.