Prostate Segmentation and Tumor Detection from MR Images Using Latent Features
Prashant Ramesh Kharote, Manoj S. Sankhe, Deepak P Patkar · 2019
A latent feature learning technique is used to segment prostate from Magnetic Resonance Imaging. An automatic segmentation of prostate is very important and challenging step in prostate cancer investigation. Accurate prostate segmentation becomes difficult due to unclear prostate boundaries as well as shape deviation among patients. Hence to tackle these difficulties latent features were extracted from MR images by deploying deep learning technique. Auto encoders are stacked together to refine the derived features. Prostate likelihood map is derived for new prostate MR image by processing these refined features. The final prostate segmentation is achieved by deploying active contour model. We used K-nearest neighbour classifier to locate cancerous region voxel-wise .The achievement of presented technique is massively tested on the dataset that includes 218 T2-weighted prostate MR images with tumors. The precision and mean absolute surface distance (MASD) are calculated to assess the performance of our method by considering manual groundtruth delineated by experienced radiologist. The average precision obtained in our study is 88.7% ±4.6%, and MASD is 1.93 mm. The accuracy and AUC obtained by K-nearest neighbour classifier of 88.2% and 0.89. The primary outcome reveals that the latent features are more proficient in prostate segmentation from MRI.