Predicting Perceived Reporting Complexity of Abdominopelvic Computed Tomography With Deep Learning

The purpose of this pilot study was to examine human and automated estimates of reporting complexity for computed tomography (CT) studies of the abdomen and pelvis.

A total of 1019 CT studies were reviewed and categorized into 3 complexity categories by 3 abdominal radiologists, and the majority classification was used as ground truth. Studies were randomized into a training set of 498 studies and a test set of 521 studies. A 2-stage neural network model was trained on the training set; the first-stage image-level classifier produces image embeddings that are used in the second-stage sequential model to provide a study-level prediction.

All 3 human reviewers agreed on ratings for 470 of the 1019 studies (46%); at least 2 of the 3 reviewers agreed on ratings for 1010 studies (99%). After training, the neural network model predicted complexity labels that agreed with the radiologist consensus rating on 55% of the studies; 90% of the incorrect predicted categories were errors where the predicted category differed from the consensus rating by one level of complexity.

There is moderate interrater agreement in radiologist-perceived reporting complexity for CT studies of the abdomen and pelvis. Automated prediction of reporting complexity in radiology studies may be a useful adjunct to radiology practice analytics.