Lesion Detection And Quality Assessment Of CT Image Based On Model Observer

Nowadays the modern information technology pushes the rapid development of medical imaging technology, which leads to many new imaging techniques and image processing methods. Medical image has become the important basis of clinical medical diagnosis and treatment. CT imaging has gotten wide application in medical field because of its high accuracy, low price and noninvasion in clinical examination. Applications of medical image always require that medical images have high quality, so there is always the requirement to find reasonable way to assess medical CT image quality, and this is of great importance to the researches aiming to improving medical image quality, and the algorithm test&optimization in medical image processing.Considering that the clinical radiologists are the ultimate users of the medical image, the most reliable method for assessing medical image quality should be subjective or based on their evaluation. However, subjective medical image quality assessment is time-consuming, too expensive, poor real-time and prone to be influenced by subjective and objective factor. But traditional objective medical image quality assessments are largely based on general natural image quality evaluation methods such as Mean Square Error (MSE) and Peak Signal to Noise Ratio (PSNR). Although these intensity difference based methods are computationally simple, they often cannot satisfy the diagnostic requirement of doctors— detecting the lesions in medical images or locating the lesion positions. Hence, researchers find a more compatible way to assess the medical image quality——model observer based on task for assessing medical image. Model observer is a function whose independent variable is the medical image observed and dependent variable is a test statistic. By comparing the computed test statistic using a thresholding operation, we can judge whether there is a lesion in the tested medical images.This dissertation focuses on the CT image quality assessment based on the model observer. First, we briefly introduce the background of the medical image assessment and give a review of current researches on medical image quality assessment. After that, we introduce some mathematic models used in model observer and analyze several ways to evaluate the performance of the model observer. This thesis also addresses some relevant feature of the Human Visual System and Human Visual System Models, which in fact lay a theoretical foundation for this work.Then, based on the morphological property of lesions in abdomen CT images, we put forward a subtractive ellipse signal to simulate the lesion in CT image and redefine the model to detect the lesion in CT image. Afterwards, we study the Channelized Joint detection and estimation Observer (CJO) proposed by Lu Zhang to detect and estimate additive signals with unknown amplitude, orientation and size in MRI image. By improving this observer, we use it to detect and estimate the subtractive signals in both synthesized (Correlated Gaussian) backgrounds and real abdomen CT image background, and then evaluate its detection and estimation performance. Experiment results show that this modified CJO has good performance in the Signal Known Statistically (SKS) detection-estimation task.At last, consider a more clinically realistic but more complex situation:an abdomen CT image for test has multiple signals, and observers do not exactly know the signals’amplitude, orientation and size, nor the number of signals on each image. To solve this problem, we follow the practice as Lu Zhang did in the the Perceptually relevant Channelized Joint detection and estimation Observer (PCJO) and introduce the visual difference predictor into the modified CJO to detect the position of the subtractive lesions in abdomen CT images. Then the abdomen CT image lesions diagnosis is simulated based on this modified PCJO model. A free-response subjective experiment is also performed to validate the performance of the modified PCJO with respect to subjective assessment of clinical radiologists. Preliminary results show that by choosing a suitable threshold and a proper number of channels there is no significant difference between the detecting performance of the modified PCJO and that of doctor observer.