Comparison of machine learning models in the interpretation of barium enemas in Hirschsprung's disease.

OBJECTIVE

To assess the diagnostic accuracy of machine learning models and ChatGPT in the interpretation of barium enemas for Hirschsprung's disease (HD), and to compare performance with that of pediatric radiologists.

MATERIAL AND

METHODS

A retrospective study of the barium enemas of patients < 15 years of age managed at a tertiary institution from 2011 to 2023 was carried out. The images were used to train AI models and divided into training, validation, and test ensembles. Performance was assessed in a separate test ensemble with anonymized images, while calculating sensitivity, specificity, and ROC curves vs. final diagnosis, and it was compared with radiologists' performance.

RESULTS

266 barium enemas from 218 patients (1,439 images in total) were included. The test ensemble consisted of 54 enemas, with 11 HD positive cases. The support vector model had a sensitivity of 72.7% and a specificity of 93%. The logistic regression model had an AUC-ROC of 0.73, with better results in anteroposterior than in lateral images. When compared with retrospective radiological reports, AI models had a classification capacity similar to that of expert professionals, with a sensitivity of 81% and a specificity of 76%.

CONCLUSIONS

AI models showed potential in supporting the diagnosis of Hirschsprung's disease based on barium enemas, with a good capacity to rule out HD. This could improve diagnostic accuracy, especially in environments with little experience or limited availability of specialist radiologists. However, further studies are required to optimize clinical application.