USE-Net: incorporating Squeeze-and-Excitation blocks into U-Net for prostate zonal segmentation of multi-institutional MRI datasets (1904.08254v2)

Abstract: Prostate cancer is the most common malignant tumors in men but prostate Magnetic Resonance Imaging (MRI) analysis remains challenging. Besides whole prostate gland segmentation, the capability to differentiate between the blurry boundary of the Central Gland (CG) and Peripheral Zone (PZ) can lead to differential diagnosis, since tumor's frequency and severity differ in these regions. To tackle the prostate zonal segmentation task, we propose a novel Convolutional Neural Network (CNN), called USE-Net, which incorporates Squeeze-and-Excitation (SE) blocks into U-Net. Especially, the SE blocks are added after every Encoder (Enc USE-Net) or Encoder-Decoder block (Enc-Dec USE-Net). This study evaluates the generalization ability of CNN-based architectures on three T2-weighted MRI datasets, each one consisting of a different number of patients and heterogeneous image characteristics, collected by different institutions. The following mixed scheme is used for training/testing: (i) training on either each individual dataset or multiple prostate MRI datasets and (ii) testing on all three datasets with all possible training/testing combinations. USE-Net is compared against three state-of-the-art CNN-based architectures (i.e., U-Net, pix2pix, and Mixed-Scale Dense Network), along with a semi-automatic continuous max-flow model. The results show that training on the union of the datasets generally outperforms training on each dataset separately, allowing for both intra-/cross-dataset generalization. Enc USE-Net shows good overall generalization under any training condition, while Enc-Dec USE-Net remarkably outperforms the other methods when trained on all datasets. These findings reveal that the SE blocks' adaptive feature recalibration provides excellent cross-dataset generalization when testing is performed on samples of the datasets used during training.

• The paper introduces a novel CNN architecture that integrates SE blocks with U-Net to improve prostate zonal segmentation accuracy.
• It demonstrates enhanced intra- and cross-dataset generalization by training on pooled multi-institutional T2-weighted MRI datasets evaluated using Dice metrics.
• The study highlights the importance of adaptive feature recalibration in addressing anatomical variability and inconsistent MRI representations.

Analysis of USE-Net for Prostate Zonal Segmentation

The paper presents USE-Net, a novel Convolutional Neural Network (CNN) architecture designed to enhance prostate zonal segmentation in multi-institutional Magnetic Resonance Imaging (MRI) datasets. USE-Net innovatively incorporates Squeeze-and-Excitation (SE) blocks into the U-Net framework to boost performance through adaptive feature recalibration, optimizing the generalization capabilities across varied datasets.

Objective and Approach

The paper targets the segmentation of the Central Gland (CG) and Peripheral Zone (PZ) within prostate MRI scans, crucial for differential diagnosis due to varying tumor prevalence and severity in these zones. Traditional methods have struggled with the blurry boundaries between the CG and PZ, and the challenge is compounded by inter-subject anatomical variability and inconsistent MRI representations across institutions.

To tackle these hurdles, the authors propose incorporating SE blocks into U-Net, examining two variants: Enc USE-Net (SE blocks after every encoder) and Enc-Dec USE-Net (SE blocks after every encoder-decoder block). These architectures were evaluated on three distinct T2-weighted MRI datasets from different institutions.

Experimental Setup

The experimentation involved evaluating USE-Net against three CNN architectures (U-Net, pix2pix, and Mixed-Scale Dense Network) and a semi-automatic continuous max-flow model over various training and testing conditions, encompassing intra- and cross-dataset scenarios. The generalization ability was tested by training on individual datasets and the union of multiple datasets, then testing across all configurations. Specifically, the paper employed Dice Similarity Coefficient (DSC) and other metrics to assess segmentation accuracy.

Key Outcomes

The results demonstrate that USE-Net variants, particularly Enc-Dec USE-Net, generally outperform other methods. This is most pronounced when trained on pooled multi-institutional datasets, reflecting enhanced intra-/cross-dataset generalization capabilities. Notably, Enc-Dec USE-Net shows robust performance when trained on the combination of all datasets, owing this improvement to the SE blocks' ability to recalibrate features adaptively, emphasizing even weaker but informative signal channels peculiar to each dataset. Conversely, the traditional U-Net and other baseline models underperform under multi-dataset conditions, indicating USE-Net's superior ability to leverage diverse data sources for model training.

The paper indicates the potential of training CNNs on multi-institutional datasets, suggesting that combining training strategies with SE blocks could be indispensable for drawing out superior model performance in practical applications.

Implications and Future Directions

The implications of this work are significant both theoretically and practically. On a theoretical level, it underscores the importance of adaptive feature recalibration in the context of biomedical image segmentation. Practically, the paper suggests that carefully designed CNNs like USE-Net can overcome traditional barriers in medical imaging tasks, potentially streamlining and enhancing diagnostic capabilities in clinical practice.

Future research might extend this work by focusing on incorporating three-dimensional spatial information to refine segmentations and exploring domain adaptation techniques through mechanisms like transfer learning. Additionally, the application of Generative Adversarial Networks (GANs) and Variational Auto-Encoders (VAEs) could offer promising pathways to further improve cross-dataset generalization and model robustness. Overall, USE-Net marks a considerable step forward in the field of medical imaging, providing a solid foundation for subsequent innovations in the area.