Vision transformer-based uncertainty quantification for triaging skin lesions: a probabilistic framework for automated biopsy recommendation.

Introduction: Automated skin lesion detection systems based on deep learning have been found to have significant potential in melanoma triage, but have not been widely used in clinical practice because they use visual features to make predictions (without patient history) and make definitive predictions with false overconfidence. This paper introduces a new probabilistic model, Evidential Deep Learning and Vision Transformer architecture, that produces safer biopsy suggestions through uncertainty quantification. Methods: The model was trained and tested using the ISIC composite database (10,010 records). The hierarchical Swin Transformer architecture is utilized in the proposed framework to extract complex features of the dermoscopic images. To combine multimodal data, a cross-attention system is trained, which learns the fine-grained correlations between visual features and patient clinical data (age, gender and lesion location). Lastly, the system, with the Dirichlet distribution in the decision layer, not only predicts the need to order a biopsy, but also produces a quantifiable uncertainty estimate that can be used to refer suspicious and out-of-distribution cases to a specialist. Results: The results of the evaluation indicate that the suggested framework had an accuracy of 92.398% and an area under the operating characteristic curve (AUROC) of 0.924. More importantly, the incorporation of evidential learning resulted in a reduction in the expected calibration error (ECE) by 0.031 and improved the model's sensitivity in detecting lesions requiring biopsy to 92.37%. It was also able to recognize 89.5% of low-quality or ambiguous images and tag them as in need of expert review. Discussion: The suggested multimodal and uncertainty-aware system not only enhances diagnostic accuracy by infusing clinical context, but also decreases human and systematic error in dermatology triage by offering a computational safety layer. This system is an effective step towards implementing Trustworthy AI in healthcare workflows.