MultiTask learning AI system to assist BCC diagnosis with dual explanation.

Basal cell carcinoma (BCC) accounts for 75% of all skin cancers. Currently, all major public hospitals in Spain have a dermatology care protocol that includes teledermatology. This has created an overload for hospital dermatologists, which could be alleviated with an AI tool for prioritization. Several AI systems have been proposed for this purpose, but the lack of transparent diagnostic explanations limits their clinical acceptance and implementation. This fact motivates the present study, which aims to develop an AI tool focused on detecting BCC from dermoscopic images incorporating dermatologist diagnostic criteria to enhance reliability. Specifically, the BCC diagnostic criterium is that a lesion is not considered BCC if it exhibits pigment network pattern, and that a lesion is considered BCC if it exhibits at least one of these BCC patterns: ulceration, ovoid nest, multi globules, maple-leaf, spoke wheel, arborizing telangiectasia. We analyzed 1,559 dermoscopic images collected from 60 primary care centers in Andalusia. Four dermatologists annotated the images as exhibiting or not each of the seven possible BCC patterns. As there is no established Ground Truth to determine the BCC patterns present in a lesion, we propose an Expectation-Maximization consensus algorithm to consolidate the multi-rater annotations into a unified standard reference (SR). As an additional novelty, the system incorporates the symbolic reasoning of dermatologists, who base their diagnoses on BCC patterns shown in lesions. To this end, a multitask learning (MTL) system based on MobileNet-V2 was designed. This system can rapidly triage BCC and non-BCC lesions while providing clinical information justifying this classification. This system also provides GradCAM-based maps to dermatologists to improve its reliability and confidence. Three evaluations were performed on the AI system. First, a performance analysis was conducted to evaluate the AI tool’s ability to classify lesions as BCC or non-BCC. In this analysis, the model achieved 90% accuracy (precision=0.90, recall=0.89). The second evaluation analyzed whether the detected patterns agreed with dermoscopic criteria. Notably, at least one clinically relevant BCC pattern was correctly identified in 99% of BCC-positive cases, and the pigment-network negative criterion was met in 95% of non-BCC cases. A comparison of the GradCAM maps with the dermatologist’s manual delineation demonstrated strong colocalization with dermatologist-segmented regions (mean foreground density 0.57 vs. background 0.16), confirming alignment of the visual focus of experts. This work introduces the first clinically validated dual-explanation AI system that combines high-accuracy BCC detection with transparent, pattern-based explanations. This approach closes the critical gap between AI performance and clinical trust in teledermatology, positioning the system for immediate deployment in primary care. Future work will focus on determining the extent to which this dual explanation system improves dermatologists’ confidence.