[Paper Review] MultiTask Learning AI system to assist BCC diagnosis with dual explanation

Iván Matas, Carmen Serrano+4 more|arXiv (Cornell University)|Jun 27, 2024

Cutaneous Melanoma Detection and Management4 citations

TL;DR

This paper presents a multi-task deep learning AI system for Basal Cell Carcinoma (BCC) diagnosis in teledermatology, combining binary BCC/non-BCC classification with clinically inspired explanation via Grad-CAM and expert-annotated dermoscopic patterns. The model achieves 90% accuracy in classification and 99% accuracy in detecting clinically relevant BCC patterns, with Grad-CAM attention strongly aligned to dermatologist-segmented regions (mean activation 0.57 inside vs. 0.16 outside), demonstrating high interpretability and clinical relevance.

ABSTRACT

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.

Motivation & Objective

• To develop an interpretable AI tool for BCC diagnosis in teledermatology to reduce diagnostic delays and optimize specialist referrals.
• To enhance model interpretability by integrating clinical knowledge through dermatologist-annotated dermoscopic patterns.
• To evaluate the alignment between AI-generated attention maps (via Grad-CAM) and clinically relevant lesion features using quantitative metrics.
• To address class imbalance in BCC datasets through data augmentation and class-weighted training.
• To validate the clinical relevance of the model’s explanations using physician-guided analysis and probability density functions of Grad-CAM activations.

Proposed method

• A multi-task deep learning framework is trained to perform binary BCC/non-BCC classification and detect multiple dermoscopic features (e.g., pigment network, ulceration, multiglobules) simultaneously.
• The model uses Grad-CAM to generate class activation maps, which are then compared to dermatologist-segmented regions of clinical interest to assess attention alignment.
• An Expectation Maximization (EM) algorithm infers a consensus ground truth for dermoscopic features from annotations by four dermatologists, addressing the lack of established ground truth.
• Conditional probability density functions of normalized Grad-CAM values are estimated within (Fg) and outside (Bg) dermatologist-segmented regions to quantitatively evaluate attention accuracy.
• The model employs class-weighted cross-entropy loss and data augmentation to mitigate class imbalance in the dataset, particularly for underrepresented patterns like spoke wheel and maple leaf-like.
• Clinical-inspired XAI is implemented by overlaying Grad-CAM maps on dermatologist-annotated lesion patterns and evaluating the spatial correspondence using mean activation and intersection of probability distributions.

Figure 1: Illustration showing the current workflow of BCC diagnosis by teledermatology in orange and the proposed workflow in green.

Experimental results

Research questions

• RQ1 Can a multi-task deep learning model achieve high accuracy in BCC/non-BCC classification while providing clinically interpretable explanations?
• RQ2 To what extent does the model’s attention, as measured by Grad-CAM, align with dermatologist-identified regions of clinical interest in BCC lesions?
• RQ3 How accurately can the model detect specific dermoscopic features used in clinical BCC diagnosis, and how does this compare to expert consensus?
• RQ4 Does the integration of physician-annotated patterns improve the reliability and clinical relevance of the AI’s explanations?
• RQ5 What is the quantitative relationship between Grad-CAM activation intensity inside and outside clinically relevant lesion regions?

Key findings

• The BCC/non-BCC binary classification achieved 90% accuracy, with a recall of 0.89 and specificity of 0.89.
• The detection of clinically relevant BCC patterns reached 99% accuracy when evaluating lesions with at least one such pattern.
• For clinically-inspired XAI, the mean Grad-CAM activation within dermatologist-segmented regions (Fg) was 0.57, compared to 0.16 outside (Bg), indicating strong attention to relevant features.
• The intersection area between the conditional probability density functions of Grad-CAM values inside and outside the clinical region was 0.24 for correct predictions, indicating higher attention alignment in correct cases.
• Incorrect predictions showed lower mean activation in Fg (0.33 vs. 0.57) and higher intersection (0.32 vs. 0.24), suggesting poor attention to clinical features in misclassified cases.
• The model demonstrated high precision (0.97) and recall (0.94) for pigment network detection, while ulceration and multiglobules showed lower precision (0.52 and 0.32, respectively), indicating challenges with rare or ambiguous patterns.

Figure 2: BCC dermoscopic criteria and Pigment Network as a negative criterion.