Automatic Segmentation of Metastatic Livers by Means of U-Net-Based Procedures.

Simple Summary: In this work, we developed three neural networks based on the U-net architecture to automatically segment the healthy liver area, the metastatic liver area, and liver metastases in micro-CT images of mice with pancreatic ductal adenocarcinoma and liver metastases. The best network for each task was then identified by cross-validation. The results demonstrated the ability of the selected networks to segment the above areas in a manner comparable to manual segmentation, at the same time saving time and ensuring reproducibility. Therefore, despite the limited number of animals involved, our pilot study represents a first step toward the development of automated tools to support liver metastasis research in the preclinical setting. Background: The liver is one of the most common sites for the spread of pancreatic ductal adenocarcinoma (PDAC) cells, with metastases present in about 80% of patients. Clinical and preclinical studies of PDAC require quantification of the liver's metastatic burden from several acquired images, which can benefit from automatic image segmentation tools. Methods: We developed three neural networks based on U-net architecture to automatically segment the healthy liver area (HL), the metastatic liver area (MLA), and liver metastases (LM) in micro-CT images of a mouse model of PDAC with liver metastasis. Three alternative U-nets were trained for each structure to be segmented following appropriate image preprocessing and the one with the highest performance was then chosen and applied for each case. Results: Good performance was achieved, with accuracy of 92.6%, 88.6%, and 91.5%, specificity of 95.5%, 93.8%, and 99.9%, Dice of 71.6%, 74.4%, and 29.9%, and negative predicted value (NPV) of 97.9%, 91.5%, and 91.5% on the pilot validation set for the chosen HL, MLA, and LM networks, respectively. Conclusions: The networks provided good performance and advantages in terms of saving time and ensuring reproducibility. [ABSTRACT FROM AUTHOR]