(Thursday, 7/16/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room: Track 4
Purpose: While liver surgery criteria is objective (disease location with respect to lobes, estimated liver volume post resection), the steps to create these estimates are subjective based on each user. The goal of this work is to reduce this variation through a deep learning algorithm to segment the eight lobes of the liver and the gross tumor volume.
Methods: For liver lobe segmentation, CT images (slice thickness 2.5-5mm) of 60 subjects were acquired with manually segmented eight lobes of the liver under an IRB protocol. For disease segmentation, CT images (slice thickness 0.7-5mm) from 116 subjects were acquired from the publicly available Liver Tumors segmentations (LiTs) challenge. 3D fully convolutional neural networks with atrous convolutions were investigated for both tasks. Similarly, a unique loss function masking the liver in the final convolution layer enabled both models to utilize the entire image while reducing the need for future class weighting. For disease segmentation, multiple atrous convolutions with residual connections are used in lieu of pooling layers.
Contours were qualitatively rated on a scale from 1-5. Trained radiologists reviewed disease contours on twenty-one scans from our institution. For liver lobe assessment, a surgeon reviewed 7 scans acquired from our institution.
Results: For liver disease segmentation of disease greater than 10cc (n=20), the DSC mean(min-max) was 0.71(0-0.95). Physician review of the disease contours (n=21) deemed 90% to be scored 4 or better. For liver lobe segmentation, the mean(min-max) DSC for each lobe was: one:0.70(0.65-0.86), two:0.79(0.67-0.88), three:0.82(0.67-0.89), four:0.76(0.45-0.87), five:0.75(0.59-0.86), six:0.73(0.58-0.81), seven:0.74(0.64-0.79), eight:0.75(0.59-0.80). Inter-observer DSC (n=1) increased on average from 0.75 to 0.80. Surgeon review of lobe contours (n=7) deemed 71% to be scored 4 or better.
Conclusion: Our models segment liver disease and liver lobes in a rapid (<30 seconds) and consistent manner. This can better facilitate clinical decisions of potential surgical eligibility.
Funding Support, Disclosures, and Conflict of Interest: Funded in part by the Helen Black Image-Guided Fund. Funded in part by the Society of Interventional Radiology Allied Scientist Grant.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
