A preprocessing method based on 3D U-Net for abdomen segmentation

Deep learning methods have made significant progress in the field of biomedical automatic segmentation but still open to developments, especially due to the insufficient use of preprocessing methods. In this study, a pre-processing step is proposed both to improve segmentation performance and to produce faster segmentation results. In this context, it is intended to obtain the abdomen region of interest (ROI) by using 3D U-Net, which has been shown to be effective in numerous studies. The presented work involves training a 3D U-Net using the CHAOS dataset and samples from the AbdomenCT-1K dataset, comprising a training dataset with 6998 slices. Afterwards, the network was exclusively tested using samples from the AbdomenCT-1K dataset, which consists of 1311 slices, to showcase its generalizability across diverse datasets. The study systematically examined the impact of fine-tuning parameters, including k value for k-fold cross-validation (CV), batch sizes (bs), and learning rates (lr), on the overall segmentation performance. Additionally, the study extended to training 3D U-Net using distinct loss functions, specifically Dice, Focal Dice, and Focal Twersky, to evaluate their respective effects on segmentation outcomes. In various scenarios, the best Dice score recorded was 99.71 %. Using the best models obtained from classical training and CV, each data in the test dataset was evaluated for its Hausdorff Distance (HD), 95th percentile of Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD). Following the segmentation process, the abdomen ROI was identified for each 2D slice using Connected Components Analysis (CCA) and established based on the largest bounding box to mitigate information loss. After applying CCA on the predicted image from the best model, an average reduction of 33.34 % in dimensionality was achieved for the entire test dataset.