TL;DR
This paper introduces ABELE, an interpretable AI (XAI) method tailored for skin lesion classification, which generates visual explanations (exemplars and counter-exemplars) to build trust in AI diagnostic systems. A user study confirmed that these explanations enhance confidence, especially among domain experts, and latent space analysis reveals distinct visual clusters for different lesion types, aiding in differentiation.
What Problem Does This Paper Address?
The increasing use of deep learning in critical fields like healthcare, particularly for medical image analysis, raises concerns about the "black box" nature of these models. Doctors need to understand why an AI makes a certain diagnosis to trust and effectively use it. Existing XAI methods often fall short in complex medical scenarios, providing explanations that are difficult to interpret or not tailored to the specific domain. This paper tackles the challenge of making AI-driven skin lesion classification more transparent and trustworthy for medical professionals, aiming to improve diagnostic accuracy and user confidence.
Methods
The core of this work lies in the ABELE (Adversarial Black box Explainer generating Latent Exemplars) framework. ABELE is a local, model-agnostic explainer that works by:
- Black Box Classifier: A ResNet architecture is trained for skin lesion classification.
- Adversarial Autoencoders (AAE): A key component is a Progressively Growing Adversarial Autoencoder (PGAAE). This advanced AAE is designed to generate realistic synthetic images (exemplars and counter-exemplars) that mimic the distribution of the real training data. The "progressively growing" aspect allows it to handle high-resolution images by starting with low-resolution and gradually increasing complexity, mitigating issues like mode collapse.
- Explanation Generation: ABELE uses the AAE to generate a neighborhood of latent representations around the input image. It then learns a local decision tree to extract rules and counterfactuals.
- Visual Explanations: The output includes:
- Exemplars: Synthetic images classified identically to the input image.
- Counter-exemplars: Synthetic images classified differently from the input image.
- Saliency Maps: Visualizations highlighting image regions critical to the classification decision.
Datasets
- ISIC 2019 Dataset: This is the primary dataset used. It contains dermoscopic images of skin lesions categorized into nine diagnostic classes: Melanoma (MEL), Melanocytic Nevus (NV), Basal Cell Carcinoma (BCC), Actinic Keratosis (AK), Benign Keratosis (BKL), Dermatofibroma (DF), Vascular Lesion (VASC), Squamous Cell Carcinoma (SCC), and Unknown (UNK). The training set has 25,331 images, and the test set has 8,238 images. The UNK category was excluded from training.
Key Insights and Ideas
- Tailored XAI for Medical Domains: Generic XAI methods are insufficient for complex medical tasks. ABELE's approach of generating synthetic exemplars and counter-exemplars provides more intuitive and domain-specific explanations.
- Trust Through Visual Evidence: Providing visual examples of what the AI "sees" (exemplars) and what it could have seen but didn't (counter-exemplars) significantly boosts user confidence in the AI's decision.
- Latent Space as a Diagnostic Tool: Visualizing the latent space of the learned representations reveals distinct clusters for different skin lesion types. This spatial separation can help differentiate between similar-looking lesions that are often misdiagnosed by humans.
- Progressive Growth for Realistic Generation: The PGAAE architecture is crucial for generating high-fidelity synthetic images, overcoming common challenges in generative models like mode collapse, especially with complex medical data.
- Expertise Matters in Explanation Interpretation: Domain experts (doctors) benefit more from detailed explanations, while novices and laypeople are better served by simpler, visually driven ones, suggesting a need for adaptive XAI.
Key Results
- Improved Classification Accuracy: The ResNet classifier achieved a balanced multi-class accuracy of 0.838 on the ISIC test set.
- Enhanced User Confidence: A user study with 156 participants showed that ABELE explanations significantly increased confidence in the AI's classification, particularly for medical experts. Confidence increased from an average of 67.69% before explanation to 77.12% after.
- Better Saliency Maps: ABELE's saliency maps were found to be more detailed and effective (lower deletion AUC, higher insertion AUC) compared to LIME and LORE.
- Latent Space Separation: Multidimensional Scaling (MDS) visualization of the latent space showed clear separation for most lesion classes, with melanomas forming a distinct central cluster. A Random Forest classifier trained on this 2D latent space achieved 85.60% accuracy in distinguishing melanoma from benign keratosis.
- Impact of Incorrect Advice: Domain experts showed a notable decrease in confidence (14%) after encountering incorrect AI advice, highlighting the importance of reliable explanations.
Limitations and Future Work
- Dataset Limitations: The ISIC dataset, while large, may not fully represent the diversity of skin lesions, skin tones, and clinical presentations, potentially leading to biases. Future work should expand the dataset.
- Patient Comprehension: The current focus is on explaining AI to medical professionals. Developing patient-friendly explanations is a future direction.
- Real-time Interaction: The visualization module is currently a demonstrator. Developing a real-time interactive system for explanation generation would require significant effort.
- Cross-Modal Explanations: Integrating natural language processing with visual explanations to create multimodal explanations is a promising avenue.
- Novel Architectures: Exploring newer architectures like Transformers for skin lesion classification and explanation is suggested.
- Longitudinal Studies: Understanding the long-term adoption and impact of AI and XAI in dermatology practice requires longitudinal studies.
Summary updated