Source-linked AI summary
USE-Net: incorporating Squeeze-and-Excitation blocks into U-Net for prostate zonal segmentation of multi-institutional MRI datasets
Leonardo Rundo, Changhee Han, Yudai Nagano, Jin Zhang, Ryuichiro Hataya, Carmelo Militello, Andrea Tangherloni, Marco S. Nobile, Claudio Ferretti, Daniela Besozzi, Maria Carla Gilardi, Salvatore Vitabile, Giancarlo Mauri, Hideki Nakayama, Paolo Cazzaniga
TL;DR
The paper addresses challenging prostate MRI zonal segmentation, where differentiating the Central Gland and Peripheral Zone matters because tumor frequency and severity differ between regions. It proposes USE-Net, which adds Squeeze-and-Excitation blocks to U-Net, and finds that multi-dataset training improves generalization, with Enc-Dec USE-Net outperforming competitor CNN architectures when trained on all datasets.
Problem
Prostate MRI zonal segmentation is challenging, while distinguishing the Central Gland from the Peripheral Zone matters because tumor frequency and severity differ between these regions.
Method
USE-Net incorporates Squeeze-and-Excitation blocks into U-Net after every Encoder or Encoder-Decoder block.
Results
Training on the union of available datasets generally outperforms separate-dataset training, while Enc-Dec USE-Net significantly outperforms competitor CNN architectures when trained on all datasets.
Takeaways & Limitations
Enc-Dec USE-Net achieves accurate prostate zonal segmentation when trained on the union of available datasets, supporting intra- and cross-dataset generalization.
Takeaways & Limitations
Future work will refine output images and investigate domain adaptation via transfer learning for cross-dataset generalization.
Abstract
from arXiv · showhide
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.
1. Introduction
Prostate zonal segmentation is clinically important but difficult because MRI appearances vary across institutions and manual delineation is time-consuming and operator-dependent. The paper proposes USE-Net, evaluates cross- and intra-dataset generalization across three T2-weighted MRI datasets, and finds benefits from multi-dataset training and SE-enhanced U-Net variants.
- Clinical motivation: CG–PZ differentiation matters clinically because tumor frequency and severity differ between these prostate regions.PZ volume ratio can refine prostate cancer diagnosis, while CG volume ratio can help monitor prostate hyperplasia.
- Clinical motivation: Reliable automatic zonal segmentation is needed because manual delineation is time-consuming and operator-dependent.Multi-institutional generalization is also essential because datasets exhibit anatomical variability and heterogeneous imaging characteristics.
- Experimental design: The study tests training on individual or multiple datasets and evaluates all training/testing combinations across three multi-institutional T2-weighted MRI datasets.USE-Net is compared with U-Net, pix2pix, Mixed-Scale Dense Network, and a semi-automatic continuous max-flow model.
- Proposed approach: USE-Net incorporates Squeeze-and-Excitation blocks into U-Net after every encoder or encoder-decoder block.The design uses adaptive channel-wise feature recalibration to improve segmentation generalization in heterogeneous clinical scenarios.
- Main findings: The study presents the first CNN-based prostate zonal segmentation investigation using T2-weighted MRI alone.It examines all possible training/testing conditions among three datasets from different institutions.
- Main findings: Training on the union of datasets generally outperforms separate-dataset training, supporting both intra- and cross-dataset generalization.Enc USE-Net generalizes well across training conditions, while Enc-Dec USE-Net performs best when trained on all datasets.
2. Related Work
Prior prostate zonal segmentation methods include classical models, multi-modal pipelines, and CNN architectures, but their multi-institutional generalization has received limited attention. This work extends U-Net with SE blocks and evaluates generalization across three institutional MRI datasets.
- Classical methods: Earlier prostate zonal segmentation studies used classifiers with statistical shape models, deformable models, active appearance models, and level sets.These approaches mainly targeted simultaneous zonal segmentation on T2-weighted MRI.
- Alternative computational methods: Other methods used continuous max-flow optimization, evidential C-means clustering, or multimodal combinations of T2w, DWI, and contrast-enhanced T1w MRI.These methods represent alternative optimization, clustering, and multi-sequence strategies for prostate segmentation.
- CNN-based methods: CNN-based work included VGG-based detection followed by sequential whole-gland and central-gland segmentation.Deep learning methods have also been applied to automated prostate cancer detection in multi-parametric MRI.
- CNN-based methods: U-Net combines contracting context extraction with symmetric expanding localization and became a successful biomedical segmentation architecture.V-Net extended U-Net for volumetric medical image segmentation, while Attention U-Net introduced attention gates for target structures.
- CNN-based methods: MS-D Net uses dense connections among features at different scales from dilated convolutions, with fewer parameters than full CNNs.Pix2pix applies conditional adversarial neural networks to image segmentation.
- Research gap: Existing literature had not addressed generalization among multi-institutional MRI datasets, limiting clinical applicability.The authors therefore compare CNN architectures across three institutions and introduce a U-Net variant incorporating SE blocks.
3. Materials and Methods
The study evaluates prostate zonal segmentation across heterogeneous, multi-institutional T2-weighted MRI datasets using several CNN architectures, a continuous max-flow model, and extensive training/testing configurations. It proposes USE-Net variants that integrate SE blocks into U-Net to recalibrate convolutional features.
- Compared methods: The evaluation compares U-Net, pix2pix, Mixed-Scale Dense Network, USE-Net variants, and a semi-automatic continuous max-flow method.Spatial-overlap and distance-based metrics are used, and results from cross-validation rounds are averaged.
- Datasets: Three multi-institutional T2-weighted MRI datasets with heterogeneous image characteristics support intra- and cross-dataset evaluation.The images vary in matrix size, slice thickness, slice counts, scanners, and annotation delineations.
- Segmentation task: The task selectively segments the Central Gland and Peripheral Zone from the whole-gland region in axial T2-weighted slices.The Peripheral Zone is obtained by subtracting the Central Gland from the whole-gland segmentation.
- USE-Net architecture: USE-Net inserts Squeeze-and-Excitation blocks after every encoder or every encoder-decoder block of U-Net.The blocks model channel-wise dependencies, boost informative features, suppress weak ones, and are investigated for strengthened feature recalibration.
4. Experimental Results
Across heterogeneous prostate MRI datasets, mixed-dataset training generally improved intra- and cross-dataset generalization. Enc-Dec USE-Net performed especially strongly with three-dataset training, while performance varied by architecture and testing condition.
- Cross-dataset generalization: Dataset #3 generally produced lower performance when excluded from training and used as the test set, likely because its distinctive image characteristics were not learned from datasets #1 and #2.Dataset #3 also comprised the highest number of patients.
- Architecture comparison: Enc USE-Net generally matched U-Net and achieved considerably better results when trained and tested on multiple datasets.It consistently outperformed U-Net under three-dataset training/testing conditions, although not always under one- or two-dataset conditions.
- Architecture comparison: Enc-Dec USE-Net consistently and remarkably outperformed the other methods when trained on all investigated datasets.Its performance was also high for CG and PZ segmentation when trained and tested on the same dataset.
- Statistical analysis: Friedman’s tests found significant differences among approaches for both the three-dataset condition and all training/testing combinations.For the three-dataset condition, p = 0.0009 for CG and p = 1.3 · 10^-6 for PZ; across all combinations, p = 0.01 for CG and p = 1.7 · 10^-10 for PZ.
- Interpretation: SE blocks’ adaptive feature recalibration was associated with excellent generalization when training included samples from each testing dataset.This recalibration boosts informative features and suppresses weak ones.
- Method comparison: Pix2pix generalized well across different datasets, MS-D Net worked better in single-dataset scenarios, and the continuous max-flow model was strongest relative to supervised methods across different datasets.The continuous max-flow model was outperformed by supervised methods when training and testing used the same datasets.
- Training conditions: Training on multiple datasets generally outperformed training on individual datasets across testing datasets and prostate zones.Training on datasets #1 and #2 generally outperformed training on dataset #1 without losing accuracy when testing on datasets #1, #2, and #3.
5. Discussion and Conclusions
The study concludes that multi-dataset training and SE-enhanced U-Net architectures support accurate prostate zonal segmentation and intra-/cross-dataset generalization in multi-institutional MRI settings. Enc-Dec USE-Net performs particularly well when trained on multiple datasets, while future work targets richer spatial modeling and domain adaptation.
- Prostate zonal segmentation: Enc-Dec USE-Net achieves accurate prostate zonal segmentation and outperforms competing methods when trained on all datasets.The evaluated task segments the central and peripheral prostate zones in multi-institutional T2-weighted MRI.
- Cross-dataset generalization: Training on the union of available datasets improves Enc-Dec USE-Net performance relative to training on individual datasets.The comparison concerns three prostate MRI datasets collected by different institutions.
- Cross-dataset generalization: Multi-dataset training generally supports both intra-/cross-dataset generalization across heterogeneous MRI datasets.The authors therefore consider training with samples from multiple devices or protocols clinically feasible.
- Architectural implications: SE blocks provide adaptive feature recalibration within U-Net, and their combination with multi-dataset training is presented as mutually indispensable for realizing their full potential.The discussion links this combination to strong generalization in the analyzed multi-institutional setting.
- Future developments: Future work will incorporate 3D spatial information among prostate MR slices and investigate domain adaptation through transfer learning.The proposed domain-adaptation direction aims to maximize distribution similarity; GANs and VAEs are mentioned as possible solutions.