Assessment of the Generalizability of Organ Segmentation CNNs Across CT Scanner Manufacturers
Presentations
PO-GePV-I-64 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: There are many useful applications for automated CT organ segmentation methods using convolutional neural networks (CNNs). Here, we assess whether CNNs trained for organ segmentation can generalize across scanner manufacturers.
Methods: CT images of 405 patients were retrospectively acquired from Siemens Healthineers (N=186, 12 scanner models) and GE Medical Systems (N=219, 8 scanner models) scanners. 18 structures were manually contoured on each CT image (liver, spleen, lung, thyroid, kidneys, pancreas, adrenal glands, bladder, aorta, bowel, stomach, heart, salivary glands, choroid plexus, and eyes). A 3D U-net was trained on the CT images using across-scanner (train Siemens, test GE, and vice-versa), and scanner-mixed (50/50 Siemens/GE training and testing) approaches. In total, 4 U-nets were trained (2 scanner-specific, 2 scanner-mixed), each with 186 training images. Performance in the test images was quantified using Dice Similarity Coefficient (DSC). Differences in across-scanner and scanner-mixed approaches was quantified using median DSC value and Wilcoxon signed rank tests.
Results: For the 219 GE images, the scanner-mixed approach had significantly higher DSC for 12/16 structures compared to the across-scanner approach. In these 12 structures, median improvements in DSC ranged from +0.001 to +0.03. The remaining 4 structures (thyroid, adrenals, stomach, aorta) did not show significant differences in across-scanner vs scanner-mixed training. In the 186 Siemens images, the scanner-mixed showed a significantly higher DSC in 8/16 structures compared to across-scanner training, while significantly lower DSC was found in 5 structures (thyroid, adrenals, heart, pituitary glands, choroid plexus). In all structures, median DSC differences of the Siemens data ranged from -0.03 to +0.02.
Conclusion: When testing the generalizability of CNNs trained on multiple scanners from a single manufacturer, results varied by manufacturer and anatomic structure. Our results indicate that manufacturer impact on segmentation of organs was minimal, even when the DSC changes were significant.
Funding Support, Disclosures, and Conflict of Interest: All authors on this abstract are employed by AIQ Solutions
Keywords
Taxonomy
Contact Email