Advanced Approaches for Bone CT Analysis Based on Deep Learning
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Computed Tomography (CT), as a 3D imaging technique, has greatly facilitated bone analysis over the past few decades. This thesis aimed to design novel deep learning approaches to analyse human bone CT. Four works have been conducted, i.e., anatomical segmentation of foot weight-bearing cone beam CT (CBCT), instance segmentation of wrist CT, semi-supervised segmentation of bone CT, and bone health analysis via bone fracture prediction.
In the first work, we developed a multi-stage method, FootSeg, for the anatomical segmentation of foot CT. FootSeg consisted of three parts, foot preprocessing, foot region segmentation, and foot bone classification. The multi-stage framework greatly simplified the implementation of the FootSeg method and achieved both qualitatively and quantitatively remarkable results. The mean Intersection-Over-Union on the bone parts was 90.3% on the testing set. To the best of our knowledge, this was the first research of fully automatic foot anatomical segmentation from weight-bearing CBCT via deep learning methods.
The second work focused on the instance segmentation of wrist CT. A novel semi-automatic method was designed to annotate 5K wrist CT slices. The annotation workload and time have been greatly reduced. An end-to-end edge reinforced U-net segmentation model was developed and demonstrated satisfying results. To the best of our knowledge, this was the first work on wrist CT instance segmentation using deep learning methods.
The third work aimed to solve the bone segmentation problem with fewer annotation data via semi-supervised learning. A patch-shuffled data transformation method was developed, and a patch-shuffle-based semi-supervised segmentation method was proposed for bone CT segmentation. Two supervised losses and a consistent unsupervised loss were employed to utilize both the labeled and unlabeled data. The proposed method was evaluated on various bone CT datasets, and the results demonstrated superior performance.
The last work was about bone health analysis via bone fracture prediction. We collected data from three population-based cohorts and processed the unstructured raw data as a structured database for model training and evaluation. We developed a deep learning-based fracture prediction model to predict the bone fragility fracture in the next five years. Compared with the clinical index of BMD T-score and FRAX, the proposed model could identify the bones with fragility fracture within five years with higher AUC values. This was the first research using the deep learning models to identify individuals with upcoming fragility fractures using wrist CT.
Please use this identifier to cite or link to this item: