This report presents a study on the application of semi-supervised deep learning for the segmentation of liver tumors and hepatic vessels in whole-body CT scans. The study utilizes a neural network model developed from the ground up, using a semi-supervised learning pipeline. The dataset used in this investigation includes 443 CT scans of liver tumors and vessels, 303 scans that have ground truth, and 140 scans without ground truth. The dataset was obtained with specific criteria and semiautomatically segmented using the Scout application. The study demonstrates the potential of deep learning techniques in medical image analysis, particularly in the context of liver tumor and vessel segmentation. By considering local 3D patches and tracking vessels to their source, deep vision transformers can reduce misclassifications and improve the overall accuracy of segmentation results, while the proposed model achieved a Dice score of 0.67 for hepatic vessels and 0.65 for tumors. Labeling the vessels within CT data is a very difficult process. It is worth noting that in some cases the model was able to locate blood vessels that were not labeled in the original ground truth annotations. The study highlights the importance of semi-supervised learning in medical image analysis, as it allows to use large amounts of unlabeled data to improve the accuracy and robustness of the segmentation results. Overall, the study shows promising results for the application of semi-supervised deep learning in liver tumor and vessel segmentation, with potential applications in other areas of medical image analysis.
Date:
2024-03-19
Relation:
EAI/Springer Innovations in Communication and Computing. 2024 Mar 19:161-174.
