Toward Liver Tumour Segmentation in CT Scans

Abstract

Semantic medical image segmentation is the process of annotating regions of interest in medical images without the need for human input. Image segmentation is the most common medical image processing task in research in the eld, and is critically important in cancer care. Small datasets are typically a challenge for researchers who must rely on preprocessing datasets to overcome this. In the task of liver tumour segmentation, datasets can also be highly variable. This research focuses on preprocessing techniques, working with a popular generalizable semantic segmentation model, nnUNet, which generalizes to new labels of di erent anatomies by analyzing a dataset without needing to be modi ed explicitly. We considered several preprocessing techniques and improved state-of-the-art Dice Similarity Coe cient (DSC), on an open-source dataset. We used CT data from the Medical Segmentation Decathlon (MSD), a crowdsourced challenge aimed at developing and benchmarking generalizable medical imaging, as well as our own dataset of metastatic colorectal liver cancer CTs (n = 198). We used nnUNet which had the highest per task, per region, DSC across all tasks in the 2018 MSD as a baseline. Our preprocessing improved state-ofthe-art liver tumour segmentation DSC and Normalized Surface Distance (NSD) by 1.5%. Although our hierarchical segmentation process similar to conventional techniques did not improve DSC or NSD, our data grouping method which is novel to the best of our knowledge improved DSC 3.0%. All our experiments were carried out on and an NVIDIA RTX 2080Ti GPU using PyTorch. Our contributions are as follows: we demonstrate that separating the data according to its voxel spacing improved prediction; we explored a hierarchical to segment tumours in liver CT scans; we also improved performance slightly with additional preprocessing that we believe keeps true to the design principles of nnUNet. To the best of our knowledge, at the time this was written, our method of separating the dataset according to voxel spacing is novel and does not appear elsewhere in the literature. Semantic segmentation will eventually be a fast and accurate tool for imaging assessment that will improve patient-care, reduce time for diagnosis, and expedite treatment.