TL;DR
This paper introduces CHAI, a novel AI system for melanoma classification that not only achieves high accuracy but also provides interpretable evidence for its decisions. By jointly learning from disease labels and human visual similarity, CHAI generates localized "activation maps" that highlight the image regions influencing its predictions, making it more trustworthy for clinical adoption.
What Problem Does This Paper Address?
Automated systems for analyzing dermoscopic images have become remarkably accurate, sometimes even surpassing human experts. However, a major hurdle to their widespread adoption in clinical practice is the lack of transparency. These "black box" systems provide predictions without explaining why they made them, leaving clinicians hesitant to rely on them. This paper addresses the critical need for interpretable AI in medical imaging, specifically for melanoma detection, by developing a system that offers evidence to support its classifications, catering to users of all skill levels.
Methods
The core of CHAI is a deep learning framework that learns a feature embedding using Convolutional Neural Networks (CNNs) and a triplet loss function. This embedding is then used for k-Nearest Neighbors (kNN) classification. What's novel here is the hierarchical triplet logic used to train the embedding. Instead of just using disease labels, the system also learns from non-expert human similarity groupings. This means the AI learns to group images that humans, even those without medical expertise, would consider similar.
The system also generates query-result activation map pairs. These maps highlight the specific regions within both the query image and its retrieved neighbors that contributed most to the distance calculation, effectively showing what the AI focused on. This is achieved by using a global average pooling layer that allows for the generation of these localized importance maps.
Datasets
The primary dataset used is the 2017 International Skin Imaging Collaboration (ISIC) challenge dataset. This dataset contains 2000 training dermoscopic images and 600 test images. For training the similarity groupings, the authors utilized a subset of 1700 images from the ISIC Archive, annotated into 37 distinct groups by non-experts who were not given disease labels. For the "Hierarchical" model, a subset of 1000 training images were used, annotated into disease-specific similarity groups.
Key Insights and Ideas
- Bridging the "User-Gap": The paper highlights the "user-gap" in AI, where even if an AI learns to distinguish diseases well, its internal representation of similarity might not align with how humans perceive similarity. The hierarchical triplet loss directly tackles this by incorporating human similarity judgments.
- Evidence as a Trust Builder: Providing visual evidence (activation maps) alongside predictions is crucial for building trust. This allows clinicians to scrutinize the AI's reasoning and potentially identify errors or biases.
- Hierarchical Similarity is Key: The hierarchical approach to learning human similarity, where similarity is considered within disease categories, proved more effective than a non-hierarchical approach that mixed all groups. This suggests that context matters for human perception of similarity.
- Localized Focus: The models trained with human similarity measures learned to focus on more localized and relevant areas of the lesion for comparison, leading to better activation map quality.
Key Results
- The CHAI system, particularly the "Hierarchical" and "Joint" models, significantly improved classification performance compared to baselines trained solely on disease labels or standard multiclass loss.
- For k=3 neighbors, the "Disease" model achieved the highest AUC. However, for larger k values (k=20, 40), the "Hierarchical" model achieved the highest AUC, comparable to top-performing systems in the ISIC challenge.
- The "Hierarchical" and "Non-Hierarchical" models showed significant improvements in matching non-expert human similarity relevancy (REL).
- Quantitative evaluation of activation maps showed over a 10% improvement in Jaccard index compared to ground truth lesion segmentations for models trained with human similarity measures.
Limitations and Future Work
- Adoption Potential: The authors acknowledge that further work is needed to determine if the CHAI method can truly improve the adoption of AI systems in clinical practice.
- Competing with Black-Box Systems: While CHAI improves interpretability, further research is needed to ensure its classification performance can consistently compete with the best "black-box" systems.
- Passive User Interaction: The authors suggest exploring how passive user interaction (e.g., users clicking on evidence results) could be used to continuously train and improve the system's performance and relevance over time.
- Generalizability: While the ISIC dataset is large, the performance on other datasets or in different clinical settings is not discussed.
Summary updated