Computer-Vision Approach to Triaging Patient-Submitted Photos of Intestinal Ostomies

Area of Science:

• Medical Artificial Intelligence
• Digital Health
• Surgical Outcomes

Background:

• Healthcare professionals receive numerous stoma images from patients.
• Managing these images requires efficient and scalable solutions.

Purpose of the Study:

• To develop and validate an AI pipeline for automated identification and triage of intestinal ostomy photos.
• To improve the efficiency of stoma care management.

Main Methods:

• A retrospective study involving 538 ostomy photos from 191 adult patients across nine hospitals.
• Fine-tuning pre-trained neural networks (MobileNetV4, ResNet50, ViT, CLIP-ViT) using 5-fold cross-validation.
• Performance evaluated using area under the receiver operator curve (AUC), precision, recall, and F1-scores.

Main Results:

• The CLIP-ViT model demonstrated superior performance in triage, achieving a macro-AUC of 0.94 and F1-score of 0.77.
• The end-to-end AI pipeline achieved an AUC of 0.99, precision of 0.87, recall of 0.93, and F1-score of 0.89.
• AI models effectively focused on stomas for classification, as indicated by attention maps.

Conclusions:

• An AI pipeline using a vision-language model accurately detects and triages patient-submitted ostomy photos.
• Prospective evaluation is recommended to support integration into digital healthcare workflows.
• The study highlights the potential of AI in streamlining ostomy care.