MultiTask Learning AI system to assist BCC diagnosis with dual explanation

This study presents a transparent, multitask learning AI system based on MobileNet-V2 and Grad-CAM that achieves 90% accuracy in detecting Basal Cell Carcinoma from dermoscopic images while providing clinically relevant pattern-based explanations to enhance trust in teledermatology workflows.

The Gist

Imagine a busy hospital where the dermatologists (skin specialists) are drowning in work. They are receiving thousands of photos of skin spots from local doctors via a "telemedicine" system. They need to sort through these photos quickly to find the dangerous ones (Basal Cell Carcinoma, or BCC) and ignore the harmless ones. But looking at every single photo takes too long, and human eyes can get tired.

Enter the new AI Assistant described in this paper. Think of it not as a robot that just says "Yes, it's cancer" or "No, it's not," but as a super-smart, transparent teaching assistant that helps the doctor make the decision.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Black Box"

Usually, AI systems are like black boxes. You put a picture in, and a number comes out. The AI might be 90% right, but it can't tell you why. If a doctor can't see the reasoning, they don't trust it. It's like a student guessing the answer on a test without showing their work; the teacher (the doctor) won't give them credit.

2. The Solution: The "Dual-Explanation" Detective

This new AI is different. It uses a technique called Multi-Task Learning. Imagine a detective who doesn't just solve the crime but also writes a detailed report explaining exactly which clues led to the solution.

This AI does two things at once:

• Task A (The Verdict): It decides if the spot is BCC (bad) or not BCC (good).
• Task B (The Evidence): It points out specific visual patterns that prove its verdict, just like a human dermatologist would.

3. The "Rulebook" (How the AI Thinks)

The researchers didn't just let the AI guess. They taught it the actual rulebook that real doctors use.

• The "No-Go" Zone: If the spot has a "Pigment Network" (a specific net-like pattern), the AI knows: "Okay, this is likely NOT cancer."
• The "Go" Zone: If the spot has any of six specific patterns (like "Maple Leaf," "Spoke Wheel," or "Ulceration"), the AI knows: "This IS likely cancer."

The AI is trained to look for these specific patterns. If it finds one, it raises a red flag. If it finds the "No-Go" pattern, it gives a green light.

4. The "Flashlight" (Visual Proof)

To make sure the AI isn't just hallucinating, the researchers gave it a flashlight (called Grad-CAM).

• When the AI says, "I see a Maple Leaf pattern," it also lights up that exact spot on the image with a heat map.
• The researchers then compared this "AI flashlight" with the actual drawings made by human experts.
• The Result: The AI's flashlight shone exactly where the human experts were looking. It wasn't looking at the background or random noise; it was focusing on the exact same clues the doctors use.

5. The "Lightweight" Engine

Most powerful AI systems are like heavy trucks—they need massive, expensive computers to run. This AI is built on MobileNet, which is like a sleek, fuel-efficient electric car.

• It's small and fast.
• It can run on standard computers in small clinics without needing super-computers.
• This means it can be used immediately in real-world hospitals, even in places with limited resources.

The Bottom Line

This system is a trustworthy partner, not a replacement.

• Accuracy: It gets the diagnosis right about 90% of the time.
• Reliability: In 99% of cancer cases, it successfully spots at least one of the "danger signs" to justify its alarm.
• Trust: Because it shows its work (the patterns) and points to the evidence (the heat map), doctors can trust it enough to use it to prioritize their workload.

In short: This AI is like a junior doctor who is incredibly fast, knows the rulebook perfectly, and always shows their homework. It helps the senior doctors (the specialists) focus on the patients who really need help, making the whole healthcare system faster and safer.

Technical Summary

1. Problem Statement

• Clinical Context: Basal Cell Carcinoma (BCC) accounts for 75% of all skin cancers. In Spain, teledermatology protocols have created a significant workload overload for hospital dermatologists due to the high volume of referrals from primary care centers.
• The Gap: While AI systems exist for skin cancer detection, they often function as "black boxes." They provide binary classifications (BCC vs. non-BCC) without transparent diagnostic reasoning. This lack of explainability limits clinical trust and adoption.
• The Challenge: Existing Explainable AI (XAI) methods (e.g., standard Grad-CAM) often only show where the model is looking, not what clinical criteria it is using. Furthermore, there is no established "Ground Truth" for BCC dermoscopic patterns, as they rely on subjective dermatologist consensus.

2. Methodology

The authors propose a Multi-Task Learning (MTL) framework designed to integrate clinical reasoning directly into the model architecture, providing a "dual explanation" system.

A. Dataset and Annotation

• Data Source: 1,559 dermoscopic images collected from 60 primary care centers in Andalusia, Spain (2022–2023).
• Annotation Strategy: Four dermatologists annotated images for the presence/absence of seven specific dermoscopic patterns:
  1. Negative Criterion: Pigment Network (PN).
  2. Positive Criteria: Ulceration (U), Ovoid Nests (ON), Multiglobules (MG), Maple Leaf-like (ML), Spoke Wheel (SW), Arborizing Telangiectasia (AT).
• Standard Reference (SR) Inference: Due to inter-rater variability (low kappa coefficients), the authors used an Expectation-Maximization (EM) algorithm to consolidate multi-rater annotations into a unified Standard Reference (SR) for training.
• Dataset Split:
  • 1,089 images for binary classification and pattern detection (with SR labels).
  • 334 images with detailed manual segmentations of specific patterns (used for XAI validation).
  • 136 non-BCC images from the ISIC Archive (mostly nevi).

B. Model Architecture

• Backbone: MobileNet-V2, chosen for its lightweight nature (3.5M parameters) and suitability for deployment in resource-constrained primary care settings.
• Multi-Task Learning (MTL) Design: The model shares a backbone but branches into three classifier layers to perform two simultaneous tasks:
  1. Binary Classification: BCC vs. Non-BCC.
  2. Pattern Detection: Identifying the presence of the 7 specific dermoscopic patterns.
• Clinical Logic Integration: The system mimics dermatologist reasoning:
  • If Pigment Network is detected $\rightarrow$ Predict Non-BCC.
  • If any of the 6 positive patterns are detected $\rightarrow$ Predict BCC.
  • If no patterns are detected $\rightarrow$ Predict Non-BCC.
• Training Strategy:
  • Phase 1: Transfer learning from ImageNet.
  • Phase 2: Fine-tuning on dermoscopic images to adapt to domain-specific features.
  • Phase 3: Retraining for pattern detection with a very low learning rate.
  • Loss Functions: Used Focal Loss to address severe class imbalance (e.g., Spoke Wheel and Multiglobules are rare).
  • Validation: 5-fold stratified cross-validation.

C. Dual Explanation Mechanism

The system provides two layers of interpretability:

1. Symbolic Explanation: A list of detected patterns (e.g., "Ulceration present") that justifies the binary decision based on clinical criteria.
2. Visual Explanation: Grad-CAM heatmaps to show the spatial location of the model's focus, validated against expert segmentations.

3. Key Contributions

1. Clinically Validated Dual-Explanation AI: The first system to combine high-accuracy BCC detection with transparent, pattern-based reasoning that aligns with established dermatological criteria.
2. Integration of Clinical Reasoning: Unlike post-hoc XAI, the model architecture itself enforces clinical logic (e.g., Pigment Network as a negative criterion).
3. Rigorous XAI Validation: The authors did not just generate heatmaps; they quantitatively validated them against expert manual segmentations using conditional probability density functions and statistical energy analysis (Cohen's d).
4. Efficiency for Deployment: Demonstrated that a lightweight model (MobileNet-V2) achieves performance comparable to massive models (ViT-Base, ConvNext) while being deployable on standard hardware without GPUs.

4. Results

• Binary Classification Performance:
  • Accuracy: 90%
  • Precision: 0.90
  • Recall: 0.89
• Pattern Detection Performance:
  • Positive Criteria: At least one clinically relevant BCC pattern was correctly identified in 99% of BCC-positive cases.
  • Negative Criteria: The Pigment Network negative criterion was correctly identified in 95% of non-BCC cases.
  • Note: While individual minority class recall (e.g., Spoke Wheel) was lower due to data imbalance, the system successfully identified at least one pattern in nearly all positive cases, which is sufficient for clinical diagnosis.
• XAI Validation (Grad-CAM vs. Expert Segmentation):
  • Spatial Alignment: Mean foreground density (activation inside the lesion) was 0.57, compared to 0.16 in the background.
  • Statistical Significance: In 98.2% of cases, the energy difference between the foreground and background was statistically significant (Cohen's d = 1.972, indicating a very large effect size).
  • This confirms the model focuses on the same regions as human experts, rather than spurious background features.

5. Significance and Impact

• Bridging the Trust Gap: By providing explanations that match clinical criteria (patterns) and visual focus (heatmaps), the system addresses the "black box" problem, fostering trust among dermatologists.
• Workflow Optimization: The system is designed for teledermatology triage. It can prioritize cases for hospital specialists, reducing waiting times and alleviating the workload of dermatologists managing referrals from 60 primary care centers.
• Resource Efficiency: The use of MobileNet-V2 allows for immediate deployment in primary care centers with limited hardware, without requiring expensive GPU infrastructure.
• Future Direction: The authors plan to conduct reader studies to measure how this dual-explanation system impacts dermatologist confidence and decision-making in real-world scenarios.

In conclusion, this work presents a robust, clinically aligned AI tool that moves beyond simple classification to offer a transparent, pattern-based diagnostic assistant, ready for immediate integration into primary care teledermatology workflows.