Evaluation of clinical applicability of automated liver parenchyma segmentation of multi-center magnetic resonance images.

Purpose: Automated algorithms for liver parenchyma segmentation can be used to create patient-specific models (PSM) that assist clinicians in surgery planning. In this work, we analyze the clinical applicability of automated deep learning methods together with level set post-processing for liver segmentation in contrast-enhanced T1-weighted magnetic resonance images. Methods: UNet variants with/without attention gate, multiple loss functions, and level set post-processing were used in the workflow. A multi-center, multi-vendor dataset from Oslo laparoscopic versus open liver resection for colorectal liver metastasis clinical trial is used in our study. The dataset of 150 volumes is divided as 81:25:25:19 corresponding to train:validation:test:clinical evaluation respectively. We evaluate the clinical use, time to edit automated segmentation, tumor regions, boundary leakage, and over-and-under segmentations of predictions. Results: The deep learning algorithm shows a mean Dice score of 0.9696 in liver segmentation, and we also examined the potential of post-processing to improve the PSMs. The time to create clinical use segmentations of level set post-processed predictions shows a median time of 16 min which is 2 min less than deep learning inferences. The intra-observer variations between manually corrected deep learning and level set post-processed segmentations show a 3% variation in the Dice score. The clinical evaluation shows that 7 out of 19 cases of both deep learning and level set post-processed segmentations contain all required anatomy and pathology, and hence these results could be used without any manual corrections. Conclusions: The level set post-processing reduces the time to create clinical standard segmentations, and over-and-under segmentations to a certain extent. The time advantage greatly supports clinicians to spend their valuable time with patients.