Loading...
Fully Automatic Segmentation of Pelvic CT Images based on Deep Learning Methods
Ghaedi, Elnaz | 2024
0
Viewed
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 57010 (46)
- University: Sharif University of Technology
- Department: Energy Engineering
- Advisor(s): Hosseini, Abolfazl
- Abstract:
- The initial step in radiotherapy treatment planning involves delineating the clinical target volumes (CTVs) and organs at risk (OARs). However, manual contouring is a time-consuming, labor-intensive, and subjective process. This study explores the potential of utilizing convolutional neural networks (CNNs) as an automated segmentation tool and an alternative to manual delineation. 218 Computed Tomography (CT) scans were gathered from two local hospitals in Tehran and the Whole Abdominal Organ Dataset (WORD). We employed ResUNet, a variant of UNet, as well as the original UNet from the Medical Open Network for AI (MONAI), an open-source framework for Deep Learning (DL) in healthcare imaging, to segment the bladder, rectum, prostate, left femoral head, and right femoral head. The quality of both models' segmentations was compared to ground truth masks using the Dice coefficient (DC) and 95th percentile Hausdorff Distance (95HD) metrics. The average DC scores obtained by ResUNet were 0.924, 0.959, 0.940, 0.774, and 0.725 for the bladder, left femoral head, right femoral head, rectum, and prostate, respectively. For UNet, the average DC scores were 0.869, 0.941, 0.954, 0.696, and 0.681 for the same structures. Additionally, the average 95HD values obtained by ResUNet were 10.33, 13.21, 14.48, 14.38, and 10.52, while for UNet, they were 14.68, 38.20, 21.66, 19.33, and 16.72 mm for the bladder, left femoral head, right femoral head, rectum, and prostate, respectively. Due to the low contrast of CT images for soft tissues, particularly those with indistinct boundaries, and the variation in size and shape, such as with the rectum and prostate, both models struggled to distinguish them as effectively as bones and the bladder. However, the accuracy of the segmentation remains acceptable and requires minimal correction, which still demands significantly less time compared to manual delineation from scratch. Notably, ResUNet demonstrated significantly superior performance, especially with soft tissues. The study demonstrates that ResUNet outperforms the benchmark UNet model and confirms that employing CNNs and DL-based models for automatic OAR segmentation in CT images saves considerable time while maintaining acceptable accuracy
- Keywords:
- Deep Learning ; Neural Networks ; Convolutional Neural Network ; Medical Images Segmentation ; Medical Open Network for AI (MONAI)Model ; Organs At Risk (OARs) ; Automatic Segmentaion ; CT Scan
- محتواي کتاب
- view