| Summary: | 碩士 === 國立成功大學 === 電腦與通信工程研究所 === 101 === One of the significant early signs of breast cancer is the appearance of micro-calcification (MC) clusters. However, it is time-consuming and exhausting to diagnose mammograms manually due to fuzzy nature of MCs and poor contrast in images. Thus, computer aided diagnosis systems are brought to assist specialists in diagnosis as a second opinion.
The MC detection scheme proposed in the thesis is composed of four stages. Firstly, detailed texture is acquired along with noise suppression. Secondly, fractal dimension and standard deviation are both applied to find candidate regions. These candidates are then connected to obtain regions of interest (ROIs). Then optical density (OD) transformation is implemented to all ROIs to enhance contrast. Thirdly, some texture analysis methods such as Haralick’s texture features and surrounding region dependence method are employed to each ROI in two forms, one in gray-level and the other in OD. Eventually, a four-fold cross validation procedure with stepwise linear discriminant analysis is used for training and testing. The trained discriminant function is used to distinguish regions with MCs from normal ones.
An experiment is conducted with 81 cases from National Cheng Kung University Hospital. The sensitivity of the proposed system achieves 96.1% with average 7.66 false positives per image (FPs/I) and 92.8% with 3.09 FPs/I. The area under ROC curve is 0.983±0.004. The results show that the proposed system strikes a balance between sensitivity and false positives and is expected to help the radiologists reduce their workload.
|
|---|
