A Comparison of CT-Based Pancreatic Segmentation Deep Learning Models.

Pancreas segmentation accuracy at CT is critical for the identification of pancreatic pathologies and is essential for the development of imaging biomarkers. Our objective was to benchmark the performance of five high-performing pancreas segmentation models across multiple metrics stratified by scan and patient/pancreatic characteristics that may affect segmentation performance. In this retrospective study, PubMed and ArXiv searches were conducted to identify pancreas segmentation models which were then evaluated on a set of annotated imaging datasets. Results (Dice score, Hausdorff distance [HD], average surface distance [ASD]) were stratified by contrast status and quartiles of peri-pancreatic attenuation (5 mm region around pancreas). Multivariate regression was performed to identify imaging characteristics and biomarkers (n = 9) that were significantly associated with Dice score. Five pancreas segmentation models were identified: Abdomen Atlas [AAUNet, AASwin, trained on 8448 scans], TotalSegmentator [TS, 1204 scans], nnUNetv1 [MSD-nnUNet, 282 scans], and a U-Net based model for predicting diabetes [DM-UNet, 427 scans]. These were evaluated on 352 CT scans (30 females, 25 males, 297 sex unknown; age 58 ± 7 years [ ± 1 SD], 327 age unknown) from 2000-2023. Overall, TS, AAUNet, and AASwin were the best performers, Dice= 80 ± 11%, 79 ± 16%, and 77 ± 18%, respectively (pairwise Sidak test not-significantly different). AASwin and MSD-nnUNet performed worse (for all metrics) on non-contrast scans (vs contrast, P < .001). The worst performer was DM-UNet (Dice=67 ± 16%). All algorithms except TS showed lower Dice scores with increasing peri-pancreatic attenuation (P < .01). Multivariate regression showed non-contrast scans, (P < .001; MSD-nnUNet), smaller pancreatic length (P = .005, MSD-nnUNet), and height (P = .003, DM-UNet) were associated with lower Dice scores. The convolutional neural network-based models trained on a diverse set of scans performed best (TS, AAUnet, and AASwin). TS performed equivalently to AAUnet and AASwin with only 13% of the training set size (8488 vs 1204 scans). Though trained on the same dataset, a transformer network (AASwin) had poorer performance on non-contrast scans whereas its convolutional network counterpart (AAUNet) did not. This study highlights how aggregate assessment metrics of pancreatic segmentation algorithms seen in other literature are not enough to capture differential performance across common patient and scanning characteristics in clinical populations.

Informations de copyright

Published by Elsevier Inc.

Déclaration de conflit d'intérêts

Declaration of Competing Interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Abhinav Suri reports financial support was provided by National Institutes of Health and reports a relationship with Springer Nature that includes royalties. Perry Pickhardt reports relationship with General Electric Company that includes consulting or advisory, Bracco Imaging SpA that includes consulting or advisory, Zebra Technologies Corp that includes consulting or advisory, Elucent that includes equity or stocks, SHINE that includes equity or stocks. Ronald Summers reports a relationship with PingAn that includes a collaborative grant, patent royalties and/or software licenses from iCAD, Philips, ScanMed, Philips, ScanMed, PingAn, Massachusetts General Brigham, and Translation Holdings; member of the Academic Radiology, Journal of Medical Imaging, and Radiology: AI Editorial Boards. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.