Exploring Uncertainty Measures in Bayesian Deep Attentive Neural Networks for Prostate Zonal Segmentation

Yongkai Liu, Guang Yang, Melina Hosseiny, Afshin Azadikhah, Sohrab Afshari Mirak, Qi Miao, Steven S. Raman, Kyung Hyun SungIEEE Access 2020 · OpenReview

Introduction

Accurate segmentation of the prostate's peripheral zone (PZ) and transition zone (TZ) on multi-parametric MRI (mpMRI) is critical for cancer staging and biopsy planning. However, manual segmentation is labor-intensive and subject to significant inter-reader variability. While deep learning (DL) offers automation, standard deterministic models lack the ability to quantify their own confidence—a vital requirement for clinical decision support. This study introduces a spatial attentive Bayesian deep learning network designed for high-accuracy segmentation with integrated uncertainty estimation.

Methodology

The proposed architecture integrates several advanced modules to capture both local details and global anatomical context.

Spatial Attentive Bayesian Network

The model consists of three primary components:

  1. Spatial Attentive Module (SAM): This module models long-range dependencies, allowing the network to leverage global anatomical relationships (e.g., the central location of the TZ relative to the PZ).
  2. Improved ResNet50 with MFPA: A modified ResNet50 backbone is enhanced by a Multi-scaled Feature Pyramid Attention (MFPA) module, which captures multi-scale anatomical features.
  3. Bayesian Uncertainty Estimation: By employing Monte Carlo (MC) dropout during inference, the model generates multiple stochastic predictions for a single input. The entropy across these predictions is calculated to produce a pixel-wise uncertainty map.
Overview of the proposed spatial attentive Bayesian deep learning architecture, including the SAM, MFPA, and uncertainty estimation process.

Experimental Setup

The model was trained and validated using the PROSTATEX dataset (304 scans). Performance was further evaluated on an External Testing Dataset (ETD) comprising 47 scans from a separate tertiary referral center using different MRI scanners (Siemens Skyra, Prisma, and Vida). Ground truth was established by research fellows and verified by expert radiologists.

Results and Discussion

The proposed method outperformed state-of-the-art architectures, including U-Net, Attention U-Net, and DeepLabV3+.

  • Internal Testing (ITD): Achieved a Dice Similarity Coefficient (DSC) of 0.80 for the PZ and 0.89 for the TZ.
  • External Testing (ETD): Maintained high performance with a DSC of 0.79 for the PZ and 0.87 for the TZ.

Crucially, the model demonstrated strong generalizability, with no significant performance drop on the external dataset, suggesting robustness across different clinical environments.

The Role of Uncertainty

A key finding was the correlation between uncertainty and anatomical complexity. Uncertainty peaked at the junctions between the PZ, TZ, and the anterior fibromuscular stroma (AFS). These regions often exhibit "partial volume effects," where tissue boundaries are blurred. The high uncertainty in these areas suggests that the model effectively flags regions where human annotators are likely to disagree, providing a valuable "map of doubt" for clinicians.

Conclusion

By combining spatial attention with Bayesian uncertainty estimation, this model provides both high-accuracy segmentation and a measure of reliability. This dual output can help radiologists identify areas requiring closer inspection, potentially reducing errors in prostate cancer diagnosis and treatment planning.

Summary updated

Abstract

Automatic segmentation of prostatic zones on multi-parametric MRI (mpMRI) can improve the diagnostic workflow of prostate cancer. We designed a spatial attentive Bayesian deep learning network for the automatic segmentation of the peripheral zone (PZ) and transition zone (TZ) of the prostate with uncertainty estimation. The proposed method was evaluated by using internal and external independent testing datasets, and overall uncertainties of the proposed model were calculated at different prostate locations (apex, middle, and base). The study cohort included 351 MRI scans, of which 304 scans were retrieved from a de-identified publicly available datasets (PROSTATEX) and 47 scans were extracted from a large U.S. tertiary referral center (external testing dataset; ETD)). All the PZ and TZ contours were drawn by research fellows under the supervision of expert genitourinary radiologists. Within the PROSTATEX dataset, 259 and 45 patients (internal testing dataset; ITD) were used to develop and validate the model. Then, the model was tested independently using the ETD only. The segmentation performance was evaluated using the Dice Similarity Coefficient (DSC). For PZ and TZ segmentation, the proposed method achieved mean DSCs of 0.80±0.05 and 0.89± 0.04 on ITD, as well as 0.79±0.06 and 0.87±0.07 on ETD. For both PZ and TZ, there was no significant difference between ITD and ETD for the proposed method. This DL-based method enabled the accuracy of the PZ and TZ segmentation, which outperformed the state-of-art methods (DeeplabV3+, Attention U-Net, R2U-Net, USE-Net and U-Net). We observed that segmentation uncertainty peaked at the junction between PZ, TZ and AFS. Also, the overall uncertainties were highly consistent with the actual model performance between PZ and TZ at three clinically relevant locations of the prostate.
BibTeX
@inproceedings{liu2020exploring,
  author = {Yongkai Liu and Guang Yang and Melina Hosseiny and Afshin Azadikhah and Sohrab Afshari Mirak and Qi Miao and Steven S. Raman and Kyung Hyun Sung},
  title = {{Exploring Uncertainty Measures in Bayesian Deep Attentive Neural Networks for Prostate Zonal Segmentation}},
  booktitle = {IEEE Access},
  year = {2020},
  url = {https://openreview.net/forum?id=r1kptLriRh},
}
Figure