Deep Learning Binary Classification of Inflammation during Keratinocyte Carcinoma Treatment with Topical Applications of Imiquimod, 5-fluorouracil, and Retinoids

Background: Keratinocyte carcinomas (KCs) require monitoring during topical therapy treatment. Artificial intelligence could enable automated assessment of inflammation status to support telemedicine-based care.
Objective: Evaluate three deep learning approaches for binary classification of inflammation status in KCs treated with topical imiquimod, 5-fluorouracil, and retinoids (IMI/5-FU/RET).
Methods: Retrospective study analyzing 2,170 clinical images of KCs collected from 2011-2024 at a dermatology practice. Images were labeled as having "active" or "inactive" inflammation by a physician and split into training (70%), validation (20%), and testing (10%) sets. Approaches used included a convolutional neural network (CNN), pre-trained feature extractors with linear probing using ResNet50, MobileNetV2, and DenseNet121 architectures, and pre-trained feature extractors with K-nearest neighbor classification.
Results: ResNet50 with K-NN classification achieved the best overall performance on test data, with 89.7% accuracy, 87.35% recall, 90.95% precision, and 88.85% F1 score.
Limitations: The dataset consisted primarily of fair-skinned individuals from a single dermatology practice, limiting generalizability. Images sometimes contained artifacts and variability in lighting and camera parameters.
Conclusion: Pre-trained deep learning models effectively classified inflammation status in KC treatment and could enable automated monitoring of topical therapy response and support telemedicine workflows. External validation on diverse populations and clinical settings is needed before implementation.