Abstract

In this thesis, two design methodologies are proposed, and also applied in the development of 2 computer-vision systems for computation-efficient and high-quality brain-tumor detection.

The first methodology aims at developing systems to detect, by conventional image processing procedures, 3D-object locations with a pixel-wise precision. The main operations of the detection are predicting gray-level distribution of the pixels in the object region and, based the prediction result, identifying/removing regions of non-interest. As 3D inputs can be sliced into axial, coronal or sagittal slice series, the prediction/identification/removal operations are performed step-by-step to the 3 series, respectively. Each removal increases the density of the object-information, facilitating the next prediction. To comprehend the pixel distributions with their locations, a 2D histogram presentation is proposed. In the design of the brain-tumor detection system, it is used to highlight the left-right asymmetry of a brain structure. Since the asymmetry is caused by tumors and non-pathological elements, an adaptive histogram modulation method is proposed to enhance the former by attenuating the latter. The prediction/identification/removal operations transform a 3D brain image into a tumoral minimum bounding box, which is then transformed into a tumor mask using simple morphological operations. The test results, on 1251 samples, have confirmed the high quality of the prediction of the tumor data distributions and the tumor detection.

The second methodology is proposed to design CNN (convolutional neural network) systems handling a complex task of brain-tumor segmentation, i.e., classifying the pixels of a brain image into 4 classes of intra-tumoral regions and the background. The methodology is to decompose this complex task into simple subtasks and each of them is performed by a simply-configurated and independently-trained CNN. By doing so, one can optimize the use of computing power and minimize the gradient conflict in training. The 4-class classification is decomposed into 3 binary classifications. Each of them is further decomposed into 2: first locating the object region and then identifying the pixels inside the region, performed by 2 independent-and-simple modules. The overall system, requiring only 0.75M trainable parameters, has been trained/tested with BraTS datasets, and its processing quality is among the best reported recently.