MultiTask Learning AI system to assist BCC diagnosis with dual explanation

Basal cell carcinoma (BCC) accounts for about 75% of skin cancers. The adoption of teledermatology protocols in Spanish public hospitals has increased dermatologists’ workload, motivating the development of AI tools for lesion prioritization. However, limited transparency in current systems hinders clinical acceptance. This study proposes an AI system for BCC detection from dermoscopic images that integrates dermatologist diagnostic criteria based on specific dermoscopic patterns. We analyzed 1559 dermoscopic images from 60 primary care centers annotated by four dermatologists for seven BCC patterns. An Expectation-Maximization consensus algorithm was used to build a unified standard reference. A multitask learning model based on MobileNet-V2 was developed to classify lesions and identify clinically relevant patterns, supported by Grad-CAM visual explanations. The system achieved 90% accuracy in BCC classification (precision 0.90, recall 0.89). Clinically relevant BCC patterns were correctly detected in 99% of positive cases, and the pigment network exclusion criterion was satisfied in 95% of non-BCC cases. Grad-CAM maps showed strong spatial agreement with dermatologist-defined regions. The proposed system combines accurate BCC detection with transparent pattern-based explanations, helping bridge the gap between AI performance and clinical trust in teledermatology.

💡 Research Summary

The paper addresses the growing workload in Spanish public hospitals caused by the widespread adoption of teledermatology, focusing on the need for an artificial‑intelligence (AI) tool that can prioritize basal cell carcinoma (BCC) cases while providing transparent, clinically meaningful explanations. Existing AI systems achieve high classification performance but lack interpretability, limiting their acceptance by dermatologists. To bridge this gap, the authors propose a dual‑explanation AI system that (1) classifies dermoscopic images as BCC or non‑BCC and (2) simultaneously detects the seven dermoscopic patterns that dermatologists use to diagnose BCC (ulceration, ovoid nests, multiglobules, maple‑leaf, spoke‑wheel, arborizing telangiectasia, and pigment network as a negative criterion).

A dataset of 1,559 dermoscopic images was assembled from 60 primary‑care centers in Andalusia during 2022‑2023. Four board‑certified dermatologists annotated each image for the presence or absence of the seven patterns, resulting in multilabel data. Because no histopathological ground truth exists for individual dermoscopic patterns and inter‑observer agreement is modest, the authors employed an Expectation‑Maximization (EM) consensus algorithm to fuse the multiple annotations into a single “Standard Reference” (SR) used for model training. The dataset was split into three subsets: (i) 1,089 images with multilabel annotations for training the binary and pattern classifiers, (ii) 334 images with expert‑drawn segmentation masks for evaluating visual explanations, and (iii) 136 non‑BCC images from the ISIC archive to represent real‑world device heterogeneity.

The core model follows a multi‑task learning (MTL) paradigm built on MobileNet‑V2, a lightweight convolutional neural network suitable for deployment in resource‑constrained telemedicine settings. A shared feature extractor feeds two heads: an binary BCC/non‑BCC classifier and a seven‑output multilabel pattern detector. Training proceeds in three stages: (1) transfer learning from ImageNet with a learning rate of 1 × 10⁻⁵ for 100 epochs, (2) fine‑tuning of the last three MobileNet blocks at a lower learning rate, and (3) a final low‑rate fine‑tuning of the pattern head. To mitigate the pronounced class imbalance (e.g., spoke‑wheel and multiglobules are rare), focal loss is applied, and stratified 5‑fold cross‑validation ensures robust performance estimates. Regularisation includes a dropout rate of 0.3 and the AdamW optimiser.

Performance is evaluated on three fronts. First, the binary BCC classifier achieves 90 % accuracy, with a precision of 0.90 and recall of 0.89. Second, pattern detection is highly reliable: in 99 % of BCC‑positive cases the model identifies at least one clinically relevant pattern, and in 95 % of non‑BCC cases the pigment‑network negative criterion is correctly recognized. Third, Grad‑CAM visual explanations are compared with expert‑drawn masks; the average foreground density of the heatmaps is 0.57 versus 0.16 for background, indicating strong spatial concordance with dermatologist‑identified regions.

The authors argue that the dual‑explanation approach directly addresses clinicians’ need for “why” information: pattern‑level outputs provide explicit diagnostic reasoning, while Grad‑CAM maps confirm that the model’s attention aligns with the visual features that experts use. The lightweight MobileNet‑V2 backbone ensures that inference can be performed quickly enough for real‑time teledermatology triage.

Limitations include the modest size of the dataset, lack of demographic metadata (age, sex, skin type), and the focus on BCC alone without assessing differentiation from other skin cancers such as melanoma or squamous cell carcinoma. The study also relies on a single geographic region, which may affect generalizability. Future work will expand the dataset across multiple institutions, incorporate additional skin‑cancer classes, and conduct prospective trials to quantify how the system influences diagnostic confidence, referral rates, and overall workflow efficiency in teledermatology.

In summary, this work presents the first clinically validated AI system that couples high‑accuracy BCC detection with transparent, pattern‑based explanations, offering a practical solution for AI‑assisted triage in teledermatology and paving the way for broader clinical adoption.