Inter-domain transfer learning for Semantic segmentation of Peripheral Zone of Prostate

Segmentation of the prostate zones, including the peripheral zone (PZ) and the transition zone (TZ), from magnetic resonance imaging (MRI) is a key part of the image based prostate cancer diagnosis and management. Among various segmentation methods, deep learning based semantic segmentation methods have drawn increasing attention in recent years due to their effectiveness compared with manual or traditional atlas-based methods. However, this method possesses two common limitations: class imbalance and sub-optimal feature extraction from noisy images. In this paper, we propose an Inter-Domain Transfer Learning (IDTL) method to alleviate these limitations in order to segment the PZ of the prostate with a better accuracy. We improve the generalization ability of the deep learning model by pretraining it to segment the entire prostate first before training to segment the PZ. Results show that the deep learning model, pretrained on segmenting the entire prostate first, results in a greater accuracy of PZ segmentation. The accuracy improves further when an additional prostate MRI dataset is included in the pretraining. This improvement in accuracy is true for each fold of the 10-fold cross validation and each patient of the main dataset, which evidences the effectiveness of the proposed method. We report cases where the deep learning model yields poor segmentation accuracy to identify causes and solutions in the limitations of our proposed model.