Comprehensive autoencoder for prostate recognition on MR images

Automatic recognition of anatomical structures is an essential prerequisite in computer aided diagnoses (CAD) such as tissue segmentation, physiological signal measurement and disease classification. However, insufficient color and speckle information in medical images pose challenges to the recognition of anatomical structures. Such challenges are evident with prostate recognition on magnetic resonance (MR) images and thus remain an open problem, although prostate cancer is an important problem that are attracting increasing interests in medical imaging. In this study, we propose an automatic approach for prostate recognition on MR images. Firstly, compared to existing works which integrate autoencoder with a specific type of classifier, we let autoencoder itself serve as a classifier and therefore lessening the impact from irregular and complex background found in prostate recognition. Secondly, an image energy minimization scheme with consideration of the coherence information from neighboring pixels is proposed to improve the recognition results with clear boundary appearance. We evaluate our method in comparison with three widely applied classifiers and the phase of atlas-based seeds-selection in prostate segmentation on a public prostate database. Our experiment results demonstrate significant superiority of our method in terms of both precision and F-measure.