University of New Mexico Health Sciences Center Albuquerque, NM
Award: ACG Presidential Poster Award
Evelyn Inga, MD1, Abhishek Patel, MD1, Pooja Viswanath, DO2, Ahmed Telbany, MD2 1University of New Mexico Health Sciences Center, Albuquerque, NM; 2University of New Mexico, Albuquerque, NM Introduction: Patient Bristol Stool Form Scale (BSFS) self-assessment is often unreliable. We trained and validated a smartphone-compatible convolutional neural network (CNN) to classify stool images by BSFS type, offering an objective alternative to patient diaries.
Case Description/
Methods: We obtained 208 de-identified stool photographs covering BSFS types 1 through 7. Two physicians (one internal medicine resident and one gastroenterology fellow) independently labeled every image; disagreements were resolved by consensus. Images were randomly divided into an 80 percent training set and a 20 percent validation set. The Google Teachable Machine default CNN architecture was trained for 30 epochs with a learning rate of 0.0001. The primary outcome was overall validation accuracy. Secondary outcomes included per-class precision, recall, F1 score, and comparison with physician performance on the same images. A 30-day pilot diary study compared automated classifications with patient self-reports. Discussion: The CNN achieved an overall validation accuracy of 83 percent. Per-class precision ranged from 63.6 percent for type 5 to 92.1 percent for type 1 (Figure 1). Mean sensitivity and specificity across classes were 79.4 percent and 86.0 percent, respectively. Two blinded physicians achieved a mean accuracy of 79.0 percent; the difference versus the model was not statistically significant (P = 0.42). In the pilot study the CNN correctly classified 124 of 150 images (83 percent), exceeding patient diary concordance, which was 67 percent, by 16 percentage points.
A CNN trained with 208 images classified stool morphology with accuracy comparable to expert physicians, demonstrating feasibility for objective BSFS assessment in clinical practice and research. Future work will enlarge the dataset, improve performance on under-represented classes, and prospectively assess clinical impact. Institutional review board approval was not required because all images were de-identified and collected for quality improvement.
Figure: Per-class precision of the CNN on the validation dataset.
Disclosures: Evelyn Inga indicated no relevant financial relationships. Abhishek Patel indicated no relevant financial relationships. Pooja Viswanath indicated no relevant financial relationships. Ahmed Telbany indicated no relevant financial relationships.
Evelyn Inga, MD1, Abhishek Patel, MD1, Pooja Viswanath, DO2, Ahmed Telbany, MD2. P5103 - Automated Image Based Classification of Stool Morphology Using a Mobile Convolutional Neural Network, ACG 2025 Annual Scientific Meeting Abstracts. Phoenix, AZ: American College of Gastroenterology.
