Uncertainty quantification and out-of-distribution detection in skin and breast lesion diagnostics using conformal prediction

This study introduces a novel framework combining CP and out-of-distribution (OOD) detection to enhance the reliability of medical image diagnostics. Focused on skin and breast lesion classification, the approach employs pretrained convolutional neural networks (e.g., ResNet18) and statistical confidence intervals to provide interpretable predictions. The framework uses the D7P dermatology and BreastMNIST datasets, simulating rea-lworld diagnostic scenarios where images with artifacts act as OOD examples. A meticulous data preparation strategy, including manual annotation mapping, stratified dataset splits, and systematic inclusion of artifact-based OOD cases, ensures robust evaluation. The model achieves classification accuracies of 75.28% on skin lesions and 87.78% on breast ultrasound images, with CP coverage probabilities consistently exceeding 94.8%. OOD detection incorporates latent-space embeddings and GRAM, achieving AUROC scores of 0.8339 under Gaussian noise (σ = 0.5) and near-perfect results for invert transformations. These innovations give clinicians reliable confidence measures to improve decision-making.