Detection of neonatal pneumoperitoneum on radiographs using deep multi-task learning

Comprehensive Summary

The aim of the study was to develop a deep multi-task learning AI model capable of diagnosing neonatal pneumoperitoneum using radiographs and evaluate its utility compared to clinicians of varied experience. The retrospective study developed internal data from a pediatric radiograph dataset from the University of Ulsan between January 1995 and August 2018 to select 3,861 images from children less than 3 months old. Two radiologists reviewed these images and labeled 294 cases as pneumoperitoneum and 252 cases as no pneumoperitoneum, specifically highlighting free air margins. The external data set obtained 378 images from 11 hospitals which were reviewed by three pediatric radiologists who labeled 164 cases with pneumoperitoneum and 214 cases without. The developed AI model used a shared encoder, DenseNet-169, to extract imaging features from combined classification and segmentation tasks. The reader study included four physicians, two specialists (a pediatric radiologist with 7 years experience and a neonatologist with 8 years), and two trainees (4th year residents, one pediatric and one radiology). The external dataset was split in half to form Dataset A and Dataset B. The physicians examined Dataset A with AI assistance and Dataset B without assistance, waited 5 weeks, and then examined Dataset B again, this time with AI assistance and Dataset A without assistance. This AI model outperformed conventional models that detect for pneumoperitoneum with an AUC of 0.98 compared with 0.74. Overall, AI assistance in the reader study increased accuracy 4.1%, recall 4.6%, and precision 3.7%, especially for the pediatric resident and neonatologist. The AI model performed better on internal testing over external testing with an AUC of 0.98 over 0.89 showing promise for generalization, but also indicating that the model was potentially overfitted to the training data. The assistance of this AI model has potential to aid medical practitioners of varying expertise in quick, radiologic diagnosis of pneumoperitoneum.

Outcomes and Implications

The model was trained on data that lacked supporting CT scans or ultrasound imaging to determine definitively if pneumoperitoneum was present, and was overfitted on training data. This indicates that the model will require more diverse training and the application of advanced regularization techniques before it can be implemented or tested in a real-world setting. However, the improvements to diagnosis and consensus between physicians show that this AI model has great promise as an assistance tool for physicians.