Artificial intelligence-assisted detection of adhesions on cine-MRI

Adhesive disease, which commonly occurs as a postoperative complication, is a major cause of morbidity and places a substantial burden on healthcare worldwide. Currently, laparoscopy is the only accurate diagnostic technique for abdominal adhesions, which intrinsically involves health risks including the formation of new adhesions. Non-invasive diagnostic methods with similar reliability are lacking. In recent years using cine-MRI scans of the abdomen captured during respiration has demonstrated promising performance in the diagnosis of adhesions. However, correct interpretation of cine- MRI scans requires considerable radiological expertise and this technique has not been widely adopted in clinical practice yet. In this masters thesis, the first fully-automated multistage computer-aided diagnosis (CAD) method for adhesion detection is proposed. The method exploits the phenomenon of visceral slide, a pattern of abdominal motion observed during respiration in healthy subjects. Local reduction of visceral slide is a diagnostic criterion of adhesions. Visceral slide that occurs on a cine-MRI slice is quantified using a segmentation mask generated by a deep learning model and a deformation field between cine-MRI frames obtained with an image registration algorithm. Bounding boxes of adhesions are predicted with a region growing method based on the visceral slide values. Additionally, false positives reduction driven by domain knowledge is performed. The impact of using all cine-MRI time points and different normalisation options are investigated and the hyper-parameters of the method are determined with 5-fold cross-validation. When evaluated with cross-validation, the best method configuration yields detection sensitivity of 0:61 and 0:73 at 1 and 2 false positives per slice along with 0:53 AUC in slice-level diagnosis. On the held-out test set, a slightly different configuration is top-performing and achieves detection sensitivity of 0:7 and 0:91 at 1 and 1:89 false positive per slice along with 0:78 slice-level AUC, which indicates the promising potential of the core idea of the method.