Abstract
While artificial intelligence (AI) has significantly improved skin disease classification, most models rely on the International Classification of Diseases, Tenth Revision (ICD-10), which was designed for clinical record-keeping rather than biological feature recognition. This study introduces a new taxonomy based on cytology and pathology ("Taxonomy 2") and evaluates its predictive performance using a Deep Residual Learning method. Our results demonstrate that this biologically aligned taxonomy achieves superior accuracy, F1-scores, and generalization compared to the standard ICD-10 system, suggesting high potential for clinical application in dermatology.
The Problem with Traditional Classification
In dermatology, diagnostic accuracy for general practitioners often ranges between 48% and 77% due to the visual similarity of over 1,500 distinct conditions. While computer vision models offer a solution, they are typically trained on ICD-10 categories. However, ICD-10 was built for human billing and record-keeping, not for the cellular and organic features that deep learning models actually extract from images. This misalignment can lead to models that struggle to distinguish between diseases that look different clinically but share similar cellular origins.
Methods: A Cytological Approach
To address this, researchers developed Taxonomy 2, a six-level hierarchical tree structure built from the bottom up based on cellular similarities. The system is organized around three primary root nodes:
- Keratinogenic diseases (originating in keratinocytes)
- Melanogenic diseases (originating in melanocytes)
- Non-KC and non-MC diseases (related to other cell types)
The study utilized a Deep Residual Learning (ResNet-50) model pre-trained on ImageNet. To compare the systems, the researchers established four "Projects": Project 1 served as the ICD-10 baseline, while Projects 2-4 represented increasingly granular levels of the new cytological taxonomy.
Dataset and Training
The researchers aggregated a pool of 56,571 images from eight public databases. To ensure a rigorous head-to-head comparison, they curated a subset of 1,160 images where diseases were clearly identifiable in both the ICD-10 and the new cytological system. For each project, two-thirds of the images were used for training, and one-third served as the test group.
Key Results
The most significant findings emerged from Project 4, the most detailed level of the new taxonomy. Taxonomy 2 demonstrated a remarkable ability to generalize to unseen data compared to the ICD-10 baseline.
Key performance highlights include:
- F1-Score Improvement: In the test group, the new taxonomy achieved an F1-score of 91.88%, significantly outperforming the ICD-10 baseline of 74.43% ($P = 0.005$).
- Reliability: The Kappa scores were consistently higher for Taxonomy 2, indicating more reliable predictions that account for chance.
- Stability: While ICD-10 performance dropped sharply when moving from training to testing data, Taxonomy 2 remained stable, proving its superior "extension" to unknown diagnostic data.
Discussion and Conclusion
The success of Taxonomy 2 suggests that AI performs best when labels align with underlying biology. By grouping diseases by cellular origin, the visual features the AI extracts—such as texture and cell structure—align more closely with the categorical labels. This hierarchical flexibility allows for high accuracy even at broad levels of diagnosis.
While future work should incorporate histopathological and dermoscopic images to further refine the model, this study proves that a cytologically-based taxonomy provides a more robust foundation for AI-assisted dermatology than current clinical standards.
Summary updated