Density map estimation with convolutional neural networks to count radiopaque markers on colonic transit studies.

BACKGROUND

A radiopaque marker study measures colonic transit time for work-up of primary constipation. It requires the patient to ingest multiple tiny radiopaque markers, which the radiologist must count manually on follow-up abdominal radiographs. Counting these markers is tedious but cognitively simple.

OBJECTIVE

To develop a convolutional neural network (CNN) capable of counting the number of radiopaque markers on abdominal radiographs.

MATERIALS AND

METHODS

The image archive at a large tertiary children's hospital was searched to identify abdominal radiographs performed in children for the indication of a radiopaque marker study. To establish the ground truth, a radiologist manually labeled the coordinates of the radiopaque markers in each radiograph and thereby generated a density map for that radiograph. A CNN was trained to estimate this density map from its corresponding abdominal radiograph. Spatially integrating the output density map provided an estimate of the number of markers in the radiograph. To assess model accuracy, mean absolute error and root mean square error were calculated.

RESULTS

The study cohort consisted of 436 radiographs (mean number of markers per radiograph: 34). This cohort was randomly divided into training, validation and testing sets consisting of 306, 65 and 65 radiographs, respectively. Based on the testing set, the CNN accurately estimated the number of markers in each radiograph with mean absolute error=2.6 markers and root mean square error=3.9 markers.

CONCLUSION

The proposed CNN generated promising results in counting the number of radiopaque markers on abdominal radiographs and offers the potential of automating the interpretation of colonic transit studies.