Advancing Segmentation and Classification Methods in Magnetic Resonance Imaging via Artificial Intelligence
Skip to main content
Open Access Publications from the University of California


UCLA Electronic Theses and Dissertations bannerUCLA

Advancing Segmentation and Classification Methods in Magnetic Resonance Imaging via Artificial Intelligence


Magnetic Resonance Imaging (MRI) is a non-invasive imaging technology that produces detailed anatomical images, which has provided a giant leap forward in medical diagnosis. MRI segmentation and classification play an essential role in disease assessment and detection, volume measurement, and biopsy. Methods relying on feature engineering have traditionally been used to perform MRI segmentation and classification and usually produce sub-optimal results. With the fast growth of artificial intelligence, deep learning has achieved great success and outperformed these traditional methods. However, current deep learning models, especially in prostate MRI segmentation and classification, may provide an insufficient representative power, and lack prediction uncertainty information and prior knowledge such as cancer heterogeneity, and usually require large-scale and generalizability evaluation. Primarily with prostate MRI, this dissertation concerns several advanced deep learning-based MRI segmentation and classification methods or applications to address the above issues. The contributions of this thesis are as follows:1. A new deep learning method with feature pyramid attention to enhance multi-scaled and high-level feature extraction was developed for automated prostate zonal segmentation. The proposed method outperformed state-of-art deep learning-based prostate zonal segmentation method such as U-Net. 2. The attention mechanism improves the deep learning-based segmentation by focusing more on relevant information to the region of interest, and Bayesian statistics equips deep learning with uncertainty measurement. An attentive Bayesian deep learning network was developed for the prostate zonal segmentation with uncertainty estimation. The proposed method was superior to the first method developed above on prostate zonal segmentation. Uncertainties produced between different prostate zones at three prostate locations were consistent with the actual model performance. 3. Texture provides the prior knowledge that can quantitatively describe the tumor heterogeneity. Texture-based deep learning (Textured-DL), which can be potentially used in a small dataset due to the exploitation of tumor prior information, was proposed for the prostate cancer classification. The Textured-DL showed superior performance to the radiologist-based classification, conventional machine learning, and deep learning methods. 4. A previously developed deep learning model in the second method above, attentive Bayesian deep learning network, was evaluated for the whole prostate gland segmentation using a large patient cohort. In the qualitative evaluation, the deep learning method demonstrated acceptable or excellent segmentation quality in most cases. The deep learning method was superior to the state-of-art deep learning methods in the quantitative evaluation. 5. A previously developed deep learning model in the second method above, attentive Bayesian deep learning network, was tailored and used for the placental segmentation on longitudinal MRI to investigate the model’s generalizability for other biomedical image applications. The deep learning model can automatically segment the placenta with high accuracy. In addition, placental volume measurement with the deep learning-based and manual segmentation can be used interchangeably. In summary, the deep learning model with feature pyramid attention and attentive Bayesian deep learning method achieved superior prostate zonal segmentation performance; enriching the image prior knowledge to the deep learning enhances the prostate cancer classification; large-scale and generalizability evaluation further demonstrated the segmentation model's outstanding segmentation and generalizability abilities. Future studies will explore the prior knowledge that can enhance the segmentation performance, study the contour-based fast segmentation using graph convolution, explore the non-textured and clinical features that could enhance the classification performance, and analyze the effect of data size on the Textured-DL’s classification performance.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View