Multitask AI detects BCC with dual explanations

Per the arXiv preprint (eprint arXiv:2407.00104), the authors propose a multitask learning system to assist diagnosis of basal cell carcinoma (BCC) from dermoscopic images. The study assembled 1559 images from 60 primary care centres annotated by four dermatologists and applied an Expectation-Maximization consensus algorithm to create a unified reference standard, according to the preprint. The reported model is built around MobileNet-V2 and jointly predicts lesion-level BCC presence and seven clinically relevant dermoscopic patterns, augmented with Grad-CAM visual explanations.

Per the preprint, the training pipeline uses multitask supervision to produce both the binary BCC label and pattern-level outputs. The authors report classification performance of 90% accuracy with precision 0.90 and recall 0.89 for BCC detection (per the preprint). Pattern-detection performance is reported as correct identification in 99% of positive BCC cases, and the pigment-network exclusion criterion holds in 95% of non-BCC cases (per the preprint). The paper includes Grad-CAM maps and quantitative spatial-agreement analysis showing alignment between model saliency and dermatologist-marked regions.

Editorial analysis: Multitask approaches that surface clinically meaningful sublabels - here, dermoscopic patterns - address an important clinician-facing transparency gap that single-label classifiers often leave open. Explicit pattern outputs, when paired with localisation maps like Grad-CAM, can make model behaviour more interpretable to specialists and may simplify clinical validation endpoints compared with purely opaque predictors.

For practitioners: key next steps are external and prospective validation on independent multisite cohorts, evaluation of calibration and decision thresholds in low-prevalence screening settings, and assessment of whether pattern-level outputs change triage or biopsy decisions in practice. Researchers and deployers should also compare saliency-based explanations with alternative explanation paradigms and quantify inter-rater agreement for pattern annotations when scaling beyond the original annotation team.

The preprint demonstrates a concrete workflow that pairs multitask learning with saliency visualisations to improve interpretability for BCC detection on dermoscopic images. The results are promising but remain a preprint-stage contribution pending external validation and clinical evaluation.