Exploiting Confident Information For Weakly Supervised Prostate Segmentation Based On Image-Level Labels

Prostate segmentation on magnetic resonance images (MRI) is an important step for prostate cancer diagnosis and therapy. After the birth of deep convolution neural network (DCNN), prostate segmentation has achieved great success in supervised segmentation. However, these works are mostly based on abundant fully labeled pixel-level image data. In this work, we propose a weakly supervised prostate segmentation (WS-PS) method based on image-level labels. Although the image-level label is not sufficient for an exact prostate contour, it contains potential information which is helpful to make sure a coarse contour. This information is referred to confident information in this paper. Our WS-PS method includes two steps which are mask generation and prostate segmentation. First, the mask generation (MG) exploits a class activation maps (CAM) technique to generate a coarse probability map for MRI slices based on image-level label. These elements of the coarse map which have higher probability are considered to contain more confident information. To make use of confident information from coarse probability map, a similarity model (S-Model) is introduced to refine the coarse map. Second, the prostate segmentation (PS) uses a residual U-Net with a size constraint loss to segment prostate based on the refined mask obtained from MG. The proposed method achieves a mean Dice similarity coefficient (DSC) of 83.39% as compared to the manually delineated ground-truth. The experimental results indicate that our weakly supervised method can achieve a satisfactory segmentation on prostate MRI only with image-level labels.