Abstract
Automatic skin lesion recognition has shown to be effective in increasing access to reliable dermatology evaluation; however, most existing algorithms rely solely on images. Many diagnostic rules, including the 3-point checklist, are not considered by artificial intelligence algorithms, which comprise human knowledge and reflect the diagnosis process of human experts. In this paper, we aimed to develop a semisupervised model that can not only integrate the dermoscopic features and scoring rule from the 3-point checklist but also automate the feature-annotation process. We first trained the semisupervised model on a small, annotated data set with disease and dermoscopic feature labels and tried to improve the classification accuracy by integrating the 3-point checklist using ranking loss function. We then used a large, unlabeled data set with only disease label to learn from the trained algorithm to automatically classify skin lesions and features. After adding the 3-point checklist to our model, its performance for melanoma classification improved from a mean of 0.8867 (SD 0.0191) to 0.8943 (SD 0.0115) under 5-fold cross-validation. Our proposed semisupervised learning framework can help with the automatic diagnosis of skin disease based on its ability to detect dermoscopic features and automate the label-annotation process.
Introduction
While deep learning models, particularly Convolutional Neural Networks (CNNs), have achieved expert-level performance in skin lesion classification, they face significant hurdles regarding interpretability and data scarcity. Most AI models function as "black boxes," providing diagnoses without explaining the underlying clinical features, such as asymmetry or pigment networks. This lack of transparency can hinder clinical trust. Furthermore, training these models typically requires massive datasets of labeled images. While disease labels (e.g., melanoma vs. benign) are common, detailed annotations of specific dermoscopic features are rare due to the time and expertise required for manual labeling.
This study addresses these challenges by developing a semisupervised model that integrates human diagnostic logic—specifically the clinical "3-point checklist"—into the learning process. This approach aims to improve classification accuracy while providing an interpretable "reasoning" path that mimics human dermatologists.
Methods
The researchers developed a dual-branch semisupervised learning framework utilizing a "Mean Teacher" architecture to bridge the gap between small annotated datasets and large unlabeled ones.
Model Architecture
The framework consists of two primary components:
- Supervised Branch: A ResNet-18 backbone trained on a small, high-quality dataset (900 images) labeled with both disease type and 3-point checklist features (asymmetry, atypical pigment network, and blue-white structures).
- Ranking Loss: A ranking loss function was implemented to incorporate human logic, ensuring the model learns that the presence of these three features correlates with a higher probability of melanoma.
- Semisupervised Branch (Mean Teacher): To leverage a larger unlabeled dataset (9,839 images), the authors used a Mean Teacher approach. A "Student" model learns from the data, while a "Teacher" model (an exponential moving average of the student's weights) provides stable targets, allowing the model to "self-annotate" features on unlabeled data.
Data and Preprocessing
The study utilized images from the International Skin Imaging Collaboration (ISIC) archive. The labeled set included 900 dermoscopic images from ISIC 2019, while the unlabeled set comprised 9,839 images primarily from ISIC 2020. Images underwent rigorous preprocessing, including cropping, resizing, color constancy (using the "Shades of Gray" algorithm), and Contrast-Limited Adaptive Histogram Equalization (CLAHE) to standardize variations in lighting and skin tone.
Results
The integration of human diagnostic rules significantly enhanced model performance across both classification and feature detection tasks.
Classification and Feature Detection
Adding the 3-point checklist logic improved melanoma classification accuracy from a mean of 0.8867 (SD 0.0191) to 0.8943 (SD 0.0115) in 5-fold cross-validation. On a 400-image unlabeled test set, the model achieved an AUC of 0.9848, performing competitively with top-tier models on the ISIC 2020 leaderboard.
The model also demonstrated strong performance in identifying specific components of the 3-point checklist:
- Asymmetry: 0.80 accuracy (AUC 0.8380)
- Atypical Network: 0.89 accuracy (AUC 0.9036)
- Blue-white structures: 0.76 accuracy (AUC 0.8444)
Interpretability and Automation
The semisupervised framework successfully learned to detect dermoscopic features on unlabeled images, effectively reducing the "expert bottleneck" in medical AI development. By predicting checklist scores alongside the diagnosis, the model provides a transparent output that clinicians can verify. Interestingly, the model achieved high accuracy on "atypical networks," a feature that human annotators often find ambiguous, suggesting AI could help standardize clinical interpretations.
Discussion and Conclusion
This research demonstrates that knowledge integration is more effective than relying solely on raw pixel patterns. By "teaching" the model clinical logic through ranking loss, the researchers improved both accuracy and interpretability. The framework successfully combines semantic knowledge with computer algorithms to arrive at more accurate and interpretable diagnostic results.
However, limitations remain. The dataset primarily consisted of images from White skin, necessitating future training on more diverse skin tones to ensure equitable performance. Additionally, clinical image "noise" such as hair and shadows continues to impact performance. Future work will focus on incorporating more complex diagnostic rules (e.g., the 7-point checklist or ABCD rule) and exploring advanced ensemble architectures. Ultimately, this semisupervised framework offers a scalable path toward more reliable AI tools for dermatology.
Summary updated