Room: AAPM ePoster Library
Purpose: to the guideline for modified response evaluation criteria in solid tumors assessment for hepatocellular carcinoma (HCC), the one of important factors in decision making of HCC treatment strategies is the longest viable tumor diameter, which should be correctly and automatically measured on contrast-enhanced arterial phase computed tomography (CT) images to reduce inter- and intra-observer variabilities. Convolutional-neural-network (CNN) with lung cancer CT image-based transfer learning may be promising for segmentation of HCC regions towing to the larger number of lung cancer patients than that of HCC patients. In this study, we investigated CT-based CNN segmentation of HCC regions with transfer learning based on lung cancer data.
Methods: cases with poorly differentiated HCC, who received treatments based on tumor size, underlying liver disease and functional status of the patient, were selected from HCC patients. A deep learning architecture was a tensor-flow-based open-source CNNs (NiftyNet) for researches in medical imaging. The CNN model pre-trained with lung cancer CT images was retrained as an HCC-CNN segmentation model to segment HCC regions using CT images in training datasets. An average Dice’s similarity coefficient (DSC) and Hausdorff distance (HD) were employed for evaluation of the segmentation accuracy based on a 5-fold cross-validation test. The DSC denotes the degree of region similarity between HCC regions annotated by a radiologist and the estimated regions. The HD is defined as the distance that measures how far two subsets of a metric space are from each other.
Results: proposed segmentation model achieved the average DSC and HD of 0.792 ± 0.06 and 2.42 ± 1.14 mm, whereas the segmentation model without pre-training produced the average DSC and HD of 0.714 ± 0.10 and 3.28 ± 2.20 mm, respectively.
Conclusion: proposed approach with lung-cancer-based transfer learning showed the potential to automatically delineate the HCC regions on CT images.