Deep-Learning-Based Semantic Segmentation of Spinal Vertebrae for Automatic Human Body Composition Analysis
This is the documentation of the GitHub repository https://github.com/sagerpascal/mse-vt2. It was created by Pascal Sager as his second project thesis during his MSc studies. The project thesis is supervised by Prof. Dr. Thilo Stadelmann, Director of the Centre for Artificial Intelligence at ZHAW and Dr. Felice Burn, researcher at the Canton Hospital Aarau.
Project Description
Fine-tuned AI-driven software tools allow an automated analysis of digital images and play a highly relevant role in different industries, especially in healthcare. For instance, CT images can provide very accurate information about structural anatomy, morphology, quantitative and qualitative composition of body parts. It was found that body compositions correlate with clinical phenotype and certain diseases and even might influence patient prognosis. Therefore, applying computer vision technology for body composition analysis may contribute to higher safety, efficiency and quality during the patient journey.
Existing systems for body composition analysis usually segment parts of the upper body in a supervised fashion. This requires a lot of labeled data (e.g., 3D segmentation masks of CT scans). If the body composition can be predicted with a system that does not require masks and only specific axial slices, the labeling process, which is currently a bottleneck, can be significantly simplified.
The goal of this project thesis is to develop a deep learning model that (a) detects the position and angle of all vertebrae of the spine and (b) identifies at least the four vertebrae Th4, Th12, L3, and L4. For further analysis, slices which are orthogonal to the center of the four detected vertebrae have to be extracted. Afterwards, it is investigated if the body composition, clinical information or events can be predicted based on the pixel values (Hounsfeld Unit = HU) of these slices.
This experimentally developed and fully automated analysis tool may quantify the body compositions in a thoracic and abdominal CT imaging dataset with reliable correlation to specific clinical distribution patterns, which has further relevance in patient management and outcome.
Specific tasks:
- Survey literature in the addressed field of research
- Familiarize with required technologies such as DICOM, 3D image processing and medical vision applications
- Select baseline approach from the literature and adapt it to the given problem space
- Set up development environment and train (improve & evaluate) neural networks iteratively
- Write a scientific report with focus on motivation, methods, argumentation, and novel results (paper style, 10 pages double-column format)
- Optional: Publish a scientific paper with ZHAW and KSA researchers (if circumstances allow)