PhD Oral Exam - Yanming Sun, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Brain tumor detection, or brain tumor segmentation in terms of image processing, is a challenging task by computer vision. To achieve an acceptable processing quality, one tends to build complex computing structures, requiring enormous computing and data resources. The objective of the work presented in this thesis is to develop brain tumor segmentation systems with the emphasis on a high computation efficiency, i.e., achieving a good processing quality at a very low computation cost to enable an easy implementation and a wide applicability of the systems. To this end, two different methodologies are proposed, and also applied in the development of 2 brain tumor detection systems, respectively.

The first design methodology proposed in this thesis aims at developing systems capable to detect object locations, sizes and shapes in a 3D image, by conventional image processing procedures, without dependency on data resources. The main operations of the detection are 1) predicting gray level distribution of the pixels in the object region and 2) based the prediction result, identifying/removing regions of non-interest. The prediction and identification/removal are interleaved each other and performed iteratively. Each removal increases the density of the object information in the remaining part of the 3D image, facilitating the prediction and identification/removal in the following iteration. In the design of the system for whole tumor detection, as each input 3D brain image can be sliced into series of axial, coronal or sagittal slices, the prediction/identification/removal operations are performed in 3 iterations to the 3 series of slices, respectively. To comprehend the distributions of the gray levels of the pixels with their locations, a 2D histogram presentation is proposed. In this design, it is used to highlight the left-right asymmetry of a brain structure, in terms of gray level distribution. As such an asymmetry is related to the presence of tumors and to non-pathological causes, a novel adaptive histogram modulation method is applied to enhance the former and to attenuate the latter. By the 3-iterations, the input 3D brain image is cropped into a minimum 3D bounding box covering the tumor region, and it is then transformed into a tumor mask by means of simple morphological operations. The system has been tested extensively with the samples of more than 1000 patient cases from BraTS dataset. The test results have confirmed the high quality of the prediction of the tumor data distributions and the detection of the tumor locations & shapes.

To further detect precisely the intra-tumoral regions, i.e., classifying all the pixels inside the tumor regions into 3 classes, namely edema (ED), non-enhancing/necrotic core (NET) and enhancing tumor (ET), one needs a system performing a very large number of filtering operations to undertake such a complex task. To this end, another design methodology is proposed targeting high-quality and low-cost convolutional neural network (CNN) systems for object segmentations. It is to decompose a complex task into several simple subtasks in such a way that each of them can be performed by a simple CNN module configurated and trained independently. By doing so, one can optimize the use of computing power by custom-designing the processing structure, on one hand, and minimize the gradient conflict in training, on the other hand. For the design of the CNN system, the task of classifying the pixels into one of the 4 classes, i.e., the 3 tumoral classes and the background, is decomposed into 3 binary classifications, each of which is performed in 2 steps: first locating the object region and then identifying the pixels inside the region. Thus, the proposed CNN system is made to have 3 subsystems and each consists of 2 independent and simple modules. One generates low-resolution location maps and the other the final masks delineating its designated tumor regions. The 2 modules are made to complement each other to achieve an overall good performance at a low computation cost. The CNN system has, in total, 0.75 M trainable parameters, which is a tiny fraction of that in the other systems for the same task. It has been trained and tested with BraTS 2019 ~ 2023 datasets, and its processing quality is among the best reported recently.

This research work provides, to the topic area of computer vision for medical image processing, with the design methodologies to develop object detection/segmentation systems of high computation efficiency. It demonstrates that a high processing quality is achievable without need for high computing power. Hopefully, our work will have a positive effect in the area of object detection/recognition and let everybody pay more attention to the computation efficiency. In this way, even with limited available computation and data resources, more developments in designing efficient processing systems can be expected in this important area.