Enhanced Pelvic CT Segmentation via Deep Learning: A Study on Loss Function Effects

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiotherapy Planning

Background:

• Manual segmentation of organs at risk (OARs) in pelvic CT images for radiotherapy planning is time-consuming and prone to inter-observer variability.
• Accurate OAR delineation is critical for effective radiotherapy planning, minimizing dose to healthy tissues while maximizing tumor coverage.

Purpose of the Study:

• To evaluate the efficacy of various convolutional neural network (CNN) architectures for automated OAR segmentation in pelvic CT images.
• To compare the performance of U-Net, ResU-Net, SegResNet, and Attention U-Net models against expert manual segmentations.
• To investigate the impact of different loss functions on segmentation accuracy for OARs including the bladder, prostate, rectum, and femoral heads.

Main Methods:

• Implementation and comparison of U-Net, ResU-Net, SegResNet, and Attention U-Net models using the MONAI framework.
• Training and validation on pelvic CT datasets comprising 240 patients for prostate segmentation and 220 patients for other OARs.
• Performance evaluation using Dice Similarity Coefficient (DSC), Jaccard Index (JI), and 95th percentile Hausdorff Distance (95thHD).

Main Results:

• SegResNet demonstrated superior performance across all OARs, achieving high DSC values (e.g., 0.951 for bladder, 0.829 for prostate, 0.979 for left FH).
• Attention U-Net also showed strong results, particularly for bladder and rectum segmentation.
• Dice loss function provided optimal or equivalent performance for SegResNet across OARs, with DiceCE slightly improving prostate segmentation (DSC=0.845).

Conclusions:

• Advanced CNNs, particularly SegResNet, offer a reliable and efficient automated solution for OAR segmentation in pelvic CT.
• Optimized CNN models and loss functions can significantly enhance the precision and efficiency of radiotherapy planning.
• Automated segmentation holds promise for reducing manual workload and improving consistency in clinical radiotherapy workflows.