EP190 - Deep Learning-Based Automated Detection and Classification of Stoma Complications in Clinical Photographs: A Single-Center Validation Study

Purpose/Background: The number of patients living with a stoma is increasing, yet a substantial proportion experience complications such as peristomal skin problems, mucocutaneous separation, and prolapse. Timely recognition and management are frequently constrained by limited access to specialized ostomy care and counseling, as well as associated costs. Recent advances in deep learning have enabled automated image analysis with performance comparable to that of human experts. This study aimed to develop and validate a deep learning-based image analysis model capable of classifying stoma complications from clinical photographs, with the goal of supporting patient self-management and facilitating timely evaluation and intervention.

Methods/Interventions: We retrospectively collected 722 clinical stoma images from a tertiary care center between 2008 and 2024. Images were annotated by three board-certified colorectal surgeons, who first classified them as normal or abnormal and subsequently assigned specific complication labels to abnormal cases. Model development and evaluation were conducted using stratified K-fold cross-validation, with the dataset divided into K folds and training and validation sets rotated across iterations. Baseline performance of three models - a decision tree, a conventional convolutional neural network (CNN), and EfficientNet-B0 - was compared using accuracy, area under the receiver operating characteristic curve (AUC), specificity, sensitivity, precision, and F1-score. To mitigate shortcut learning, preprocessing strategies included stoma localization with bounding box labeling, centering, image resizing, and normalization. Additionally, data augmentation (rotation, shift, flip) was applied to improve training for underrepresented samples. Following these steps, EfficientNet-B0 was re-evaluated to assess the impact of preprocessing and augmentation on classification performance.

Results/Outcomes: Among the three models evaluated, EfficientNet-B0 demonstrated the highest overall performance, achieving an accuracy of 0.7210 ± 0.0396, AUC of 0.7953 ± 0.0138, specificity of 0.8571 ± 0.0112, and F1-score of 0.7259 ± 0.0661. Both CNN and decision tree models showed lower performance in most metrics, with the exception that CNN demonstrated higher sensitivity (0.8534 ± 0.1003) than EfficientNet-B0 (0.7143 ± 0.0921). Performance improved substantially after applying the proposed preprocessing procedure. EfficientNet-B0 achieved an accuracy of 0.8158 ± 0.0256 and AUC of 0.8947 ± 0.0278, reflecting notable gains compared with the baseline. Further enhancement was observed after incorporating data augmentation, with final accuracy and AUC reaching 0.8696 ± 0.0254 and 0.9445 ± 0.0232, respectively. Precision and F1-score also improved from 0.8128 ± 0.0483 and 0.8336 ± 0.0220 after preprocessing to 0.8850 ± 0.0407 and 0.8757 ± 0.0237 after augmentation.

Conclusion/Discussion: The proposed EfficientNet-B0 model demonstrates high accuracy and robustness in the automated classification of normal versus abnormal stoma in clinical photographs. This approach has the potential to support patient self-assessment of stoma conditions. However, further studies are warranted to develop models capable of subclassifying abnormal cases into specific complication categories.

Performance Comparison of EfficientNet-B0 Model Before and After Preprocessing and Data Augmentation
Performance metrics of the EfficientNet-B0 model for stoma image classification before preprocessing (Baseline), after preprocessing, and after data augmentation. Values are presented as mean ± 95% confidence interval (CI).