MKNet-family architectures for auto-segmentation of the residual pancreas after pancreatic resection: a deep learning...

Accurate interpretation of CT scans after pancreatic resection is crucial for detecting abnormalities, including postoperative complications and cancer recurrence. This study investigates the feasibility and clinical utility of a novel MKNet-family deep learning architecture for auto-segmentation of the residual pancreas on postoperative CT imaging, in comparison to previous approaches.

Novel MKNet, MSKNet and MAKNet architectures were developed. Two datasets were used: the National Institutes of Health (NIH) dataset, comprising 82 annotated normal preoperative CT scans, and the IMPACT Consortium dataset (NCT06055010; https://github.com/IMPACTconsortium/IMPACT ), comprising 81 annotated postoperative CT scans obtained < 4 weeks after pancreatectomy. Performance was assessed by Hausdorff Distance (HD), 95th-percentile-HD (HD95) and Normalized Surface Distance (NSD), and secondarily by Dice Similarity Coefficient (DSC), and compared with self-implemented existing models for preoperative pancreas auto-segmentation. Qualitative evaluation was conducted by ten abdominal radiologists.

In the postoperative setting, the MAKNet architecture showed the best performance, with an HD and HD95 of 17.3 ± 11.2 mm and 11.5 ± 10.2 mm, respectively. DSC (64.9 ± 14.8%) and NSD (27.2 ± 8.2%) were comparable to the Attention-U-Net (DSC 66.0 ± 13.8%; NSD 27.8 ± 8.4%). Clinical evaluation indicated that the MKNet-family accurately defined the postoperative pancreas (i.e., requiring minimal or no modifications) in 64 of 81 segmentations (79%).

This study demonstrates the effectiveness of novel MKNet-family architectures to accurately segment the residual pancreas on postoperative CT imaging over previous approaches. This advances the state-of-the-art in pancreas auto-segmentation and may be beneficial for medical application and education, acceleration of data annotation, and future research.