3D Liver Segmentation from CT-Scan Images

Segmenting the liver from volumetric radiographs is a time-intensive task. Machine learning can be utilized to automate this process when precision is not the primary goal. The development of machine learning models is mainly constrained by GPU memory. To preserve spatial localization in segmentation, processing the entire volumetric image at once is ideal, but this approach is memory-intensive. Sophisticated models can extract more detailed features, but may require dividing the volumetric image into smaller segments to fit within memory constraints. Conversely, simpler models, while less memory-intensive, might not capture all necessary features for accurate segmentation. These three requirements – precision, memory usage, and feature extraction – have their own trade-offs and thus become the focus of this research. In this paper, a simple U-Net model is employed to process the entire volumetric image, thereby preserving spatial localization. This non-complex model is enhanced by a special technique, Organ-Centric Resampling, introduced for data preprocessing. Independent Random Fields Deformation is also applied to increase the variety of the generated image.