Digital stool color screening for biliary atresia—Comparing a deep learning approach to the analogue stool color card

OBJECTIVES: Early diagnosis of Biliary atresia (BA) improves native liver and overall survival. Pale stools are a characteristic symptom and might be detected by deep learning image recognition. We tested convolutional neural networks (CNNs) to classify pictures of pale and healthy stools and compare the results with stool color card (SCC) assessment of health care professionals (HCP) and non-health care professionals (NHCP). METHODS: One thousand six hundred and sixteen stool photos with diagnostic labels including BA and healthy controls were graded as pale or normal according to the SCC. A subset of photos (33 BA and 241 healthy) was presented to HCP and NHCP for SCC assessment to evaluate accuracy and interobserver reliability. For CNN model selection and training all images were split into training, validation, and test sets. Different visualization strategies and secondary data augmentation were used for explainability and estimating robustness. RESULTS: In SCC assessment, sensitivity by NHCP trended lower than HCP (86.4% vs. 97.3%), while there were no differences in specificity. Intra-cohort agreement on classification of BA stools (pale vs. normal) was also lower for NHCP than for HCP (36.4% vs. 78.8%), while no significant difference was observed in healthy stools. The selected CNN (ResNet18) correctly classified 97.5% of pale and 98.5% of normal labeled stools, also showing robust results in explainability methods. CONCLUSIONS: The high accuracy of CNN shows the potential for a digital stool color screening. In contrast, NHCP assessment of pale stools tends to be inconsistent, and there is high interobserver variability among HCP and NHCP.