Automatic Detection and Segmentation of Liver Tumors in Computed Tomography Images: Methods and Limitations

Abstract

Liver tumor segmentation in computed tomography images is considered a difficult task, especially for highly diverse datasets. This is demonstrated by top-ranking results of the Liver Tumor Segmentation Challenge (LiTS) achieving ~70% dice. To improve upon these results, it is important to identify sources of limitations. In this work, we developed a tumor segmentation method following automatic liver segmentation and conducted a detailed limitation analysis study. Using LiTS dataset, tumor segmentation results performed comparable to state-of-the-art literature, achieving overall dice of 71%. Tumor detection accuracy reached ~83%. We have found that segmentation’s upper limit dice can reach ~77% if all false-positives were removed. Grouping by tumor sizes, larger tumors tend to have better segmentation, reaching a maximum approximated dice limit of 82.29% for tumors greater than 20,000 voxels. Medium and small tumor groups had an upper dice limit of 78.75% and 63.52% respectively. The tumor dice for true-positives was comparable for ideal (manual) vs. automatically segmented liver, reflecting a well-trained organ segmentation. We conclude that the segmentation of very small tumors with size values < 100 voxels is especially challenging where the system can be hyper-sensitive to consider local noise artifacts as possible tumors. The results of this work provide better insight about segmentation system limitations to enable for better false-positive removal development strategies. Removing suspected tumor regions less than 100 voxels eliminates ~80% of the total false-positives and therefore, may be an important step for clinical application of automated liver tumor detection and segmentation.