AI-driven diagnosis of mpox using deep learning models

Towards Transparent Decisions: CNN Ensemble with XAI-Driven Interpretations

Real-Time Edge-to-Mobile Melanoma Screening Using Sensitivity-Optimized Deep Learning and TensorRT Acceleration

This article presents a real-time edge-to-mobile melanoma image-analysis prototype combining lesion localization, sensitivity-oriented classification, ensemble fusion, and TensorRT-optimized embedded inference. Six classification architectures were evaluated across HAM10000, ISIC 2019, and ISIC 2020 using a unified training and sensitivity fine-tuning protocol. The selected ResNet-50 and MedFusionNet stack achieved AUC values of 0.985, 0.988, and 0.940, respectively, under the benchmark metadata conditions. At the frozen global threshold of θ=0.20, the corresponding sensitivities were 1.000, 0.981, and 0.948. These results are classification-stage metrics computed on full, uncropped test images; sharpness rejection and YOLO-based localization were evaluated separately. Replacing patient-specific metadata with the neutral deployment vector changed sensitivity by at most one melanoma case per benchmark dataset. On an NVIDIA Jetson Orin Nano, the complete pipeline with TensorRT FP16-optimized classifiers operated at more than 22 FPS with an end-to-end latency of 43.6 ms. The system has not been clinically validated, and the open-source implementation is provided to support reproducibility and further research.

Robust Patient-Level Evaluation of Post-Training Quantization for Edge-Deployable MPox Skin Lesion Classification

In this study, we implemented Post-Training Quantization (PTQ) for Mpox skin lesion classification using a patient-level evaluation protocol based on a stringent pre-registered experimental design. The publicly available MSLD v2.0 dataset was preprocessed at the pixel level, resulting in 755 unique original images from six diagnostic classes representing 524 patients. A stratified patient-level split (80% training, 10% validation, and 10% testing) was applied to prevent data leakage. Each experiment was repeated across three independent data splits and three training seeds to assess performance variability and robustness. A two-stage transfer learning framework based on MobileNetV2 was trained and evaluated using FP32, Float16 PTQ, and full INT8 quantization implemented through TensorFlow Lite (TFLite). The FP32 baseline achieved an accuracy of 69%, a macro F1-score of 69%, and a macro-AUC of 94.7%, with a model size of 4.27 MB. Float16 quantization maintained comparable predictive performance and model size, whereas full INT8 quantization reduced the model size to 2.59 MB, representing a 39% reduction compared with the FP32 baseline. Top-2 accuracy remained stable at approximately 89% across all precision regimes, supporting the feasibility of deployment in triage-oriented clinical settings. Variability analyses revealed that performance fluctuations were driven primarily by patient partitioning rather than reduced numerical precision, demonstrating that INT8 quantization preserves clinically relevant performance while substantially reducing computational and storage requirements for deployment in resource-constrained edge environments.

A Reproducible Deep Learning Pipeline for Augmentation-Enhanced Classification of Anemia-Related Blood-Cell Images

This study presents a reproducible end-to-end deep learning pipeline for proof-of-concept classification of anemia-related and control-like blood cell images using a publicly accessible online dataset. The workflow integrates online data download, automatic extraction, dataset restructuring, class balancing, augmentation-enhanced preprocessing, transfer-learning-based model training, visualization, and structured metric generation within a single executable framework. A balanced binary dataset was constructed to reduce the effect of severe class imbalance, and augmentation strategies, including flipping, rotation, color jitter, and affine transformation, were applied only during training. A ResNet18-based classifier showed strong convergence under a fixed random seed and single train-validation-test split. On a small held-out test subset of 32 images, all images were correctly classified. However, this result should be interpreted cautiously because the test subset was small, no cross-validation or external validation was performed, and performance on limited balanced data may overestimate generalizability. Class-distribution plots, representative sample panels, learning curves, confusion matrices, prediction files, and metric summaries support transparent inspection and reproducibility. The main contribution of this study is a reproducible engineering workflow for blood-cell image classification, rather than a claim of algorithmic novelty or clinical readiness.

Skin Lesion Classification in Low-Resource Settings Using Lightweight CNNs with Uncertainty Estimation

Dermoscopic skin lesion classification is a task of major clinical importance but is computationally expensive, making it inaccessible in resource-constrained healthcare settings. In this paper, we introduce a computationally efficient skin lesion classification framework for seven classes using EfficientNet-B0, complemented by Monte Carlo (MC) Dropout for uncertainty quantification. Our approach was trained and tested on the HAM10000 dataset containing 10,015 dermoscopic images across seven classes. To address the severe 67:1 class imbalance, we employ WeightedRandomSamplerand class-weighted cross-entropy loss as complementary corrections acting at the batch-composition level and the gradient-magnitude level respectively. By performing T=50 stochastic forward passes during inference, we decompose predictive uncertainty into aleatoric and epistemic components and apply an entropy-based referral threshold that flags uncertain predictions for specialist review. To validate spatial interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) is applied and quantitatively evaluated via Intersection over Union (IoU) against ISIC segmentation masks, yielding a mean IoU of 0.61 across all accepted predictions. Our experiments achieve a test macro AUROC of 0.9404and macro F1-score of 0.7308, with six of seven classes exceeding 70% per-class accuracy (melanocytic nevi: 69.8%). Referring the 30% most uncertain predictions to a clinician raises accepted-subset AUROC from 0.9404 to 0.9568 (+1.64%). The framework is competitive with ResNet-50 and DenseNet-121 at one-fifth the parameter count, and the only lightweight method in the comparison providing calibrated uncertainty estimates. Inference latency benchmarks on an NVIDIA Jetson Nano (edge CPU mode) are reported to contextualize deployment feasibility.

Adversarial Evaluation of Multi-Organ Lassification in Computational Pathology with CNNS and Foundation Models

In this research, we aim to evaluate the robustness of convolutional neural networks (CNNs) and foundation models, such as ConCH and UNI, in classifying different types of organs from whole-slide images (WSIs) collected from various countries, different scanners, and clinical environments. Despite the dataset's inherent diversity, our results reveal that even a small perturbation (a white-box attack) with an intensity of 0.01 significantly impacts performance. ResNet-50 experienced a 67% reduction, ConCH 19%, and UNI 9.5% in accuracy. However, the foundation models performed much better, showing greater resilience and maintaining comparatively strong predictive performance even under the same intensity. Still, randomizing the dataset does not necessarily make your model resilient or robust, emphasizing the importance of thorough generalization and stability testing before clinical deployment. These models should be rigorously tested to evaluate their generalization and reliability prior to application in real patient settings. Foundation models look promising for building reliable and general AI-based cancer diagnostic systems, but they still need to prove their reliability in real-world clinical environments before large-scale adoption.