S626 Development of an AI-Based Computer Vision Algorithm for Automated Detection of Fecal Occult Blood in Stool Images

Introduction: Rapid advances in artificial intelligence (AI) are enhancing early detection tools in medicine. Applying computer vision to fecal occult blood test (FOBT) image interpretation is promising, especially in remote or low-resource settings. We developed a computer vision algorithm to predict blood presence from FOBT stool sample photos, aiming for a fast, automated, and reliable clinical support tool. Methods: We collected 67 labeled stool images (37 negative, 30 positive by FOBT) and used transfer learning with a ResNet18 convolutional neural network (CNN) pre-trained on ImageNet. Transfer learning leverages features from large datasets to improve performance with limited data. We replaced ResNet18’s final fully connected layer for binary classification (blood presence). Model weights were fine-tuned with supervised learning using cross-entropy loss and stochastic gradient descent, enabling effective feature extraction despite the small dataset. Results: The fine-tuned model achieved 74.62% average accuracy and a 70.38% F1 score, showing balanced precision and recall. Sensitivity was 66.67%, indicating moderate detection of blood-positive samples, and specificity was 80.71%, reflecting strong exclusion of negatives. These metrics are averages from 5-fold cross-validation, enhancing robustness and reducing overfitting risk. Consistent performance across folds suggests reasonable generalization despite limited data. Results highlight that transfer learning with a lightweight CNN can extract discriminative visual features from standard stool images, providing a basis for automated fecal blood detection. Conclusion: Preliminary results suggest strong potential for computer vision as a non-invasive, automated method to detect fecal occult blood, offering a digital alternative to traditional guaiac FOBT. This could expand early colorectal cancer screening access in primary care and low-resource settings. Unlike FOBT, which needs chemical reagents and lab infrastructure, our method uses standard digital images, enhancing accessibility and cost-effectiveness. Although hyperspectral imaging shows promise, its complexity limits use. Prior AI applications in stool characterization and capsule endoscopy show the rise of image-based diagnostics in gastroenterology. Our study applies visual automation at the stool level as a practical, scalable tool to improve early colorectal cancer detection.