AI System Detects Basal Cell Carcinoma Using Clinical Patterns

A multitask learning AI system was developed to classify basal cell carcinoma from dermoscopic images while providing two forms of clinical explanation. The model was trained on 1,559 images collected from 60 primary care centers in Andalusia. Four dermatologists labeled each image for the presence or absence of seven dermoscopic patterns used in BCC diagnosis.

Researchers applied an Expectation-Maximization algorithm to combine the four sets of annotations into a single standard reference. The resulting multitask model, built on MobileNet-V2, simultaneously performs binary classification and detects individual BCC patterns.

The system flags a lesion as non-BCC if a pigment network is present. It classifies a lesion as BCC when at least one of the following patterns appears: ulceration, ovoid nest, multi globules, maple-leaf, spoke wheel, or arborizing telangiectasia.

89. In 99 percent of BCC-positive cases, the system correctly identified at least one clinically relevant pattern. In 95 percent of non-BCC cases, the pigment-network criterion was satisfied. 16. The study states that the dual-explanation approach addresses the lack of transparent reasoning that has limited clinical acceptance of prior AI tools for teledermatology triage.