In this thesis, deep belief net (DBN) is applied for detection of microcalcification (MC) clusters in digital mammograms. DBN is a relatively... Show moreIn this thesis, deep belief net (DBN) is applied for detection of microcalcification (MC) clusters in digital mammograms. DBN is a relatively new type of neural network in machine learning which can learn complex structures from data by using a deep architecture. Based on a database of 200 clinical mammograms, a 3-level unsupervised network followed by a supervised back-propagation fine-tuning classifier is trained and tested. For each location in the image, the classifier is applied to decide whether there is an MC, and subsequently all the detected MCs are grouped into clusters. Free-response receiver operating characteristic (FROC) curves are used to evaluate the performance. The performance of DBN is compared to a well-known support vector machine (SVM) detector. Experimental results show that DBN can outperform SVM. In particular, DBN can achieve a detection rate of 83% at the cost of only one false positive cluster per image. These promising results show that DBN can be adopted in the study of object detection in medical image. M.S. in Electrical Engineering, May 2012 Show less
