Size Matters: Exploring the Impact of Model Architecture and Dataset Size on Semantic Segmentation of Abdominal Wall Muscles in CT Scans

Abstract

Computed tomography (CT) scans are an excellent way of capturing the details of the abdominal muscles. Several abnormalities including hernia, tumour, and neuro-muscular diseases can be identified and estimated by CT scans, including the representation of muscle loss. Radiologists carefully segment and review every CT slice to monitor these abnormalities. This is a meticulous, painstaking, and time-consuming process. This is especially true for the abdomen, which has a lot of variations and complexities in the wall and surrounding tissues. To solve this, we annotated 451 image-mask pairs from DICOM files of the abdominal muscles of size $512 \times 512$ pixels. In this work, the trade-off between computation cost and image loss is estimated by splitting the imageries into two categories of size: $128 \times 128$ pixels and $512 \times 512$ pixels. To sustain the original image size, the authors create a resized version of the base U-Net and train it from scratch. Furthermore, two dataset categories are created to show the data-intensive quality of the models (One double in size compared to the other). To train the system using these datasets, two architectures U-Net and resized U-Net are used which are then compared through the introduction of a new loss function—“complete loss”. The resized U-Net performs better than the standard U-Net by displaying a mean complete loss of 3.52 and 10.79% for the validation and test sets, respectively. The data dependency of models is clearly shown as well.