Automatic Prostate Zonal Segmentation Using Fully Convolutional Network with Feature Pyramid Attention

Yongkai Liu, Guang Yang, Sohrab Afshari Mirak, Melina Hosseiny, Afshin Azadikhah, Xinran Zhong, Robert E. Reiter, Yeejin Lee, Steven S. Raman, Kyung Hyun Sung

Abstract

Our main objective is to develop a novel deep learning-based algorithm for automatic segmentation of prostate zone and to evaluate the proposed algorithm on an additional independent testing data in comparison with inter-reader consistency between two experts. With IRB approval and HIPAA compliance, we designed a novel convolutional neural network (CNN) for automatic segmentation of the prostatic transition zone (TZ) and peripheral zone (PZ) on T2-weighted (T2w) MRI. The total study cohort included 359 patients from two sources; 313 from a deidentified publicly available dataset (SPIE-AAPM-NCI PROSTATEX challenge) and 46 from a large U.S. tertiary referral center with 3T MRI (external testing dataset (ETD)). The TZ and PZ contours were manually annotated by research fellows, supervised by genitourinary (GU) radiologists. The model was developed using 250 patients and tested internally using the remaining 63 patients from the PROSTATEX (internal testing dataset (ITD)) and tested again (n=46) externally using the ETD. The Dice Similarity Coefficient (DSC) was used to evaluate the segmentation performance. DSCs for PZ and TZ were 0.74 and 0.86 in the ITD respectively. In the ETD, DSCs for PZ and TZ were 0.74 and 0.792, respectively. The inter-reader consistency (Expert 2 vs. Expert 1) were 0.71 (PZ) and 0.75 (TZ). This novel DL algorithm enabled automatic segmentation of PZ and TZ with high accuracy on both ITD and ETD without a performance difference for PZ and less than 10% TZ difference. In the ETD, the proposed method can be comparable to experts in the segmentation of prostate zones.

Introduction

Accurate identification of the internal architecture of the prostate is a cornerstone of modern cancer diagnosis. Prostate cancer (PCa) diagnosis relies heavily on multiparametric MRI (mpMRI) and the PI-RADS scoring system, where diagnostic criteria vary depending on the specific zone: the Peripheral Zone (PZ) or the Transition Zone (TZ). Currently, radiologists must manually segment these zones, a process that is time-consuming and highly subjective. Standard deep learning models, such as the ubiquitous U-Net, often struggle with this task because anatomical borders are frequently heterogeneous or indiscernible. This study presents a novel Fully Convolutional Network (FCN) that utilizes "Feature Pyramid Attention" to automate this process, achieving results that rival human experts.

Methods

The authors proposed a specialized FCN consisting of three primary components: a modified ResNet50 encoder, a Feature Pyramid Attention (FPA) module, and a naive decoder.

The proposed network architecture featuring a modified ResNet50 encoder and the Feature Pyramid Attention (FPA) module for multi-scale feature extraction.

To capture high-level semantic information while maintaining spatial resolution, the ResNet50 encoder was modified by removing the initial max-pooling layer and utilizing dilated bottleneck blocks. The core innovation, the FPA module, extracts features at multiple scales (3x3, 5x5, and 7x7) simultaneously. This allows the model to integrate local details with global context, which is essential when physical borders are blurry.

The study utilized 359 patient cases. The model was trained on 250 cases from the PROSTATEX dataset and tested on an internal testing dataset (ITD, n=63) and an external testing dataset (ETD, n=46) from a separate tertiary referral center to evaluate real-world generalization.

Results

Performance was evaluated using the Dice Similarity Coefficient (DSC). The proposed method demonstrated superior performance compared to the standard U-Net architecture and achieved parity with human expert consistency.

Qualitative and quantitative comparison between the proposed method and U-Net, showing higher DSC scores for both PZ and TZ segmentation.

On the internal test set, the proposed method achieved a DSC of 0.74 (PZ) and 0.86 (TZ), significantly higher than the baseline U-Net (0.69 and 0.83, respectively). In the external dataset, the model's agreement with human experts (DSC 0.74 for PZ, 0.79 for TZ) was comparable to the agreement between two human experts (0.71 for PZ, 0.75 for TZ). Notably, both the model and human experts performed better on the Transition Zone (TZ) than the Peripheral Zone (PZ), likely due to the more complex and poorly discerned boundaries of the PZ.

Discussion and Conclusion

The study highlights that multi-scale context is critical for zonal segmentation. By using the FPA module, the model successfully handles anatomical ambiguity. Furthermore, the model's stability on external data suggests it captures fundamental anatomical features rather than dataset-specific noise.

Limitations include the 2D nature of the current model, which does not account for inter-slice correlations. Additionally, performance was lower at the "base-end" (top) of the prostate, where the TZ is particularly difficult to identify. Future research should explore 3D architectures and validation across a broader range of MRI scanner manufacturers.

Summary updated

Abstract

Our main objective is to develop a novel deep learning-based algorithm for automatic segmentation of prostate zone and to evaluate the proposed algorithm on an additional independent testing data in comparison with inter-reader consistency between two experts. With IRB approval and HIPAA compliance, we designed a novel convolutional neural network (CNN) for automatic segmentation of the prostatic transition zone (TZ) and peripheral zone (PZ) on T2-weighted (T2w) MRI. The total study cohort included 359 patients from two sources; 313 from a deidentified publicly available dataset (SPIE-AAPM-NCI PROSTATEX challenge) and 46 from a large U.S. tertiary referral center with 3T MRI (external testing dataset (ETD)). The TZ and PZ contours were manually annotated by research fellows, supervised by genitourinary (GU) radiologists. The model was developed using 250 patients and tested internally using the remaining 63 patients from the PROSTATEX (internal testing dataset (ITD)) and tested again (n=46) externally using the ETD. The Dice Similarity Coefficient (DSC) was used to evaluate the segmentation performance. DSCs for PZ and TZ were 0.74 and 0.86 in the ITD respectively. In the ETD, DSCs for PZ and TZ were 0.74 and 0.792, respectively. The inter-reader consistency (Expert 2 vs. Expert 1) were 0.71 (PZ) and 0.75 (TZ). This novel DL algorithm enabled automatic segmentation of PZ and TZ with high accuracy on both ITD and ETD without a performance difference for PZ and less than 10% TZ difference. In the ETD, the proposed method can be comparable to experts in the segmentation of prostate zones.
BibTeX
@inproceedings{liu2019automatic,
  author = {Yongkai Liu and Guang Yang and Sohrab Afshari Mirak and Melina Hosseiny and Afshin Azadikhah and Xinran Zhong and Robert E. Reiter and Yeejin Lee and Steven S. Raman and Kyung Hyun Sung},
  title = {{Automatic Prostate Zonal Segmentation Using Fully Convolutional Network with Feature Pyramid Attention}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=nd7pr7AVu8},
}
Figure