Machine Learning in Medical Image Processing, Analysis and Diagnosis

Image processing and analysis, and computer aids in diagnosis, are indispensable in medical imaging. Machine leaning (ML) has become one of the most active areas of research in the medical imaging field, including medical image analysis and computer-aided diagnosis (CAD), because objects such as signals, lesions and organs in medical images may not be represented accurately by a simple equation. Thus, tasks in medical imaging essentially require “learning from examples” to determine a large number of parameters in complex models of signals, lesions and organs. Recently, as the available computational power increased dramatically, patch/pixel-based ML emerged. It uses pixel/voxel values in patches in images directly instead of features calculated from segmented regions as input information. The patch/pixel-based ML is a versatile, powerful framework that can acquire image-processing and analysis functions, including noise reduction, lesion and organ enhancement, pattern separation, segmentation and classification through training with image examples.

In this talk, patch/pixel-based MLs are overviewed to make clear a) the basic principles of the patch/pixel-based ML and b) its applications to 1) separation of bones from soft tissue in chest radiographs, 2) radiation dose reduction in CT, 3) enhancement of lesions such as lung nodules in CT, 4) CAD for lung nodule detection in chest radiography and thoracic CT, 5) distinction between benign and malignant nodules in CT, and 7) polyp detection and classification in CT colonography. The effectiveness of the above technologies was rigorously evaluated in task-based observer performance studies, and some of the technologies were translated into clinical practice.