Discriminative brain image analysis via geometrically inspired deep networks
Open Access
- Author:
- Yu, Mingzhao
- Graduate Program:
- Electrical Engineering (PHD)
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- October 03, 2023
- Committee Members:
- Madhavan Swaminathan, Program Head/Chair
Mahanth Gowda, Major Field Member
Nanyin Zhang, Outside Unit & Field Member
Mehdi Kiani, Major Field Member
Steven J. Schiff, Special Member
Vishal Monga, Chair & Dissertation Advisor - Keywords:
- neuroimaging
AI
hydrocephalus
attention regularization
brain volume estimation
ultrasound
CT
MRI
classification - Abstract:
- Over the past several decades, medical imaging techniques like magnetic resonance imaging (MRI), computed tomography (CT), and ultrasound have been extensively utilized for diagnosing and treating medical conditions. Typically, the analysis of these complex medical images falls on the shoulders of human experts, including neurosurgeons, pathologists, and radiologists. Given the intricate nature of such images and the high workload on these professionals, there is a growing need to automate the process of medical image analysis through computer-aided methods. Recent advancements in machine learning (ML) technologies have made it increasingly attainable. A primary obstacle in applying ML techniques to real-world medical imaging tasks is their reliance on both the quantity and quality of training data. In this dissertation, we tackle this challenge on two fundamental discriminative tasks in medical imaging: 1) brain image classification, which crucial for the correct diagnosis of infections or other medical conditions and 2) brain image segmentation, which plays a significant role in evaluating the treatment progress for conditions like hydrocephalus. The first part of the thesis address the infection diagnosis in CT brain images. Hydrocephalus is the leading indication for pediatric neurosurgical care worldwide. Identification of postinfectious hydrocephalus (PIH) verses non-postinfectious hydrocephalus (NPIH), as well as the pathogen involved in PIH is crucial for developing an appropriate treatment plan. In this study, we develop a geometrically inspired AI method for CT-based infection diagnosis in hydrocephalic imagery. State-of-the-art (SOTA) convolutional neural network (CNN) approaches form an attractive neural engineering solution for addressing this problem as pathogen-specific features need discovery. Yet black-box deep networks often need unrealistic abundant training data and are not easily interpreted. A novel brain attention regularizer is proposed, which encourages the CNN to put more focus inside brain regions. Our approach is then extended to a hybrid 2D/3D network that mines inter-slice information. A new strategy of regularization is also designed for enabling incorporation of geometry relation between 2D and 3D inputs. Our proposed method achieves SOTA results on a CURE Children's Hospital of Uganda dataset. Such attention regularized learning has particularly pronounced benefits in regimes where training data may be limited, thereby enhancing generalizability. The second part addresses the problem of sepsis detection using multi-view cranial ultrasound (cUS) images. The captured image set for each patient enables multi-view imagery: coronal and sagittal, with geometric overlap. To exploit it, we develop a new learning framework, called the intersection-guided Cross-view Local- and Image-level Fusion Network (CLIF-Net). Our technique employs two distinct CNN branches to extract features from both views with proposed multi-level fusion blocks. Specifically, we leverage the spatial position of these images to locate the intersecting region and enhance the semantic features using cross-attention mechanism, facilitating the acquisition of mutually beneficial and more representative features for both views. The final, enhanced features from the two views are then integrated through the image-level fusion layer, outputting predictions. We contend that our method of exploiting multi-view geometry relation enables a first of its kind, robust 3D representation. When evaluated on dataset from Mbale Regional Referral Hospital in Uganda, CLIF-Net demonstrates vastly enhanced performance, surpassing the prevailing SOTA techniques in sepsis detection. The third part focuses on the assessment of brain and cerebrospinal fluid (CSF) volumes through MRI segmentation, a critical factor in evaluating treatment interventions. Accurate volume measurements facilitate meaningful comparisons between the hydrocephalus condition in patients and normal brains, helping to identify whether brain or CSF levels deviate from standard ranges. The contribution of this section is the introduction of a hemi-brain volume estimation pipeline. This pipeline comprises a segmentation generator and a hemi-brain mask generator, both of which are deep learning-based models. Additionally, the segmentation model incorporates geometric regularization specific to the CSF region. This hemi-brain approach not only allows for effective assessment of patient progress but also minimizes the impact of artifacts in MRI scans that could lead to inaccurate volume estimation. The pipeline has been successfully applied in the Endoscopic Versus Shunt Treatment of Hydrocephalus in Infants (ESTHI) Trial. It's important to highlight that incorporating prior knowledge about geometric patterns within or between images has profoundly influenced our methods. This includes the collaborative regularization driven by the relationship between 2D and 3D data in the first part, the intersection-guided multi-view fusion approach in the second, and the total variation regularizer for CSF activation in the third. These incorporations underscore the central theme of this thesis: geometrically inspired deep networks.