Deep Learning Enables Prostate MRI Segmentation: A Large Cohort Evaluation With Inter-Rater Variability Analysis

Yongkai Liu, Qi Miao, Chuthaporn Surawech, Haoxin Zheng, Dan Nguyen, Guang Yang, Steven S. Raman, Kyunghyun SungFrontiers in Oncology 2021 · DOI

TL;DR

Researchers validated a Deep Attentive Neural Network (DANN) for automatic whole-prostate gland segmentation using a massive, continuous clinical dataset of nearly 3,700 MRI scans. The model achieved a state-of-the-art Dice score of 0.93 and received "acceptable to excellent" ratings from expert radiologists in over 96% of cases, proving its readiness for real-world clinical volume measurements and biopsy planning.

What Problem Does This Paper Address?

Accurate whole-prostate gland (WPG) segmentation is the cornerstone of prostate cancer care—it’s essential for calculating volume, planning targeted biopsies, and guiding surgical interventions. However, manual segmentation is a "bottleneck" process: it is incredibly time-consuming and prone to significant inter-rater variability (even experts don't always agree on where the boundary lies).

While many deep learning models have been proposed to automate this, most were developed and tested on small, curated datasets. This paper addresses the "clinical gap" by evaluating a high-performing architecture (DANN) on a large-scale, continuous cohort of patients to see if AI can truly handle the messy, heterogeneous reality of daily hospital data.

Methods

The study utilizes a Deep Attentive Neural Network (DANN). Unlike standard convolutional neural networks (CNNs) that treat all pixels with equal weight, DANN uses attention mechanisms to focus on specific anatomical regions of interest, which is particularly helpful for the fuzzy boundaries of the prostate.

Key Technical Innovations: * Multi-planar Workflow: The researchers implemented a two-step process. First, a model (DANN$_{cor}$) segments the prostate on coronal slices to automatically identify the "start" and "stop" points of the gland. * Efficiency Boost: By using the coronal segmentation to select only the relevant axial slices for the final segmentation (DANN$_{ax}$), they reduced total inference time by 25% compared to processing every single axial slice in a volume. * Comparison: The model was benchmarked against industry standards, specifically UNet and DeepLab v3+.

Datasets

The study is notable for its scale, using 3,698 3T MRI scans acquired between 2016 and 2020 at a single institution.

  • Training Set: 335 scans.
  • Qualitative Evaluation Set: 3,210 scans (graded by radiologists).
  • Quantitative Evaluation Set: 100 scans (measured against manual "ground truth").
  • Volume Evaluation Set: 50 scans (to test clinical utility in volume estimation).

Key Insights and Ideas

  • The "Base" is the Hardest: The model performed best in the midgland but struggled most at the "base" of the prostate. This is because the base is anatomically continuous with the bladder and seminal vesicles, creating "partial volume effects" where boundaries blur.
  • Coronal Assistance Matters: Using coronal slices isn't just about accuracy; it’s a smart engineering trick for speed. By narrowing the search space for the axial model, the system becomes much more computationally efficient.
  • Endorectal Coils (ERC) Paradox: Interestingly, while only 3 scans in the training set used an endorectal coil, the model still showed higher inter-rater agreement on ERC scans in the testing set. This suggests that the improved image contrast provided by the coil helps both the AI and the human experts, even if the AI hasn't seen many examples of it.
  • Clinical Feasibility: The high agreement between DANN-enabled volume measurements and manual measurements (96% within the 95% limits of agreement) suggests that AI can now reliably replace manual "planimetry" in a clinical setting.

Key Results

  • Quantitative Accuracy: DANN achieved a Dice Similarity Coefficient (DSC) of 0.93 ± 0.02, significantly outperforming DeepLab v3+ (0.92) and UNet (0.91) with p < 0.05.
  • Radiologist Approval: Two abdominal radiologists found the AI's work to be "acceptable or excellent" in 96% of the 3,210 testing scans.
  • Inter-Rater Reliability: The radiologists reached "substantial agreement" (κ = 0.75) when grading the AI’s performance, reinforcing that the model's output is consistently high-quality.
  • Speed: The integrated coronal-axial workflow reduced inference time from 16.4 minutes to 12.6 minutes per case (including all pre-processing steps).

Limitations and Future Work

  • Single-Center Study: All data came from one institution using Siemens scanners. The authors acknowledge that the model needs to be tested on data from different manufacturers and hospitals to ensure "generalizability."
  • Hardware Requirements: DANN is computationally heavy. The authors suggest exploring "criss-cross attention modules" in the future to reduce the GPU memory footprint.
  • Data Imbalance: The training set was light on scans using endorectal coils. Future iterations would benefit from a more diverse training set to handle various hardware configurations better.

Summary updated

Abstract

Whole-prostate gland (WPG) segmentation plays a significant role in prostate volume measurement, treatment, and biopsy planning. This study evaluated a previously developed automatic WPG segmentation, deep attentive neural network (DANN), on a large, continuous patient cohort to test its feasibility in a clinical setting. With IRB approval and HIPAA compliance, the study cohort included 3,698 3T MRI scans acquired between 2016 and 2020. In total, 335 MRI scans were used to train the model, and 3,210 and 100 were used to conduct the qualitative and quantitative evaluation of the model. In addition, the DANN-enabled prostate volume estimation was evaluated by using 50 MRI scans in comparison with manual prostate volume estimation. For qualitative evaluation, visual grading was used to evaluate the performance of WPG segmentation by two abdominal radiologists, and DANN demonstrated either acceptable or excellent performance in over 96% of the testing cohort on the WPG or each prostate sub-portion (apex, midgland, or base). Two radiologists reached a substantial agreement on WPG and midgland segmentation (<i>κ</i> = 0.75 and 0.63) and moderate agreement on apex and base segmentation (<i>κ</i> = 0.56 and 0.60). For quantitative evaluation, DANN demonstrated a dice similarity coefficient of 0.93 ± 0.02, significantly higher than other baseline methods, such as DeepLab v3+ and UNet (both p values < 0.05). For the volume measurement, 96% of the evaluation cohort achieved differences between the DANN-enabled and manual volume measurement within 95% limits of agreement. In conclusion, the study showed that the DANN achieved sufficient and consistent WPG segmentation on a large, continuous study cohort, demonstrating its great potential to serve as a tool to measure prostate volume.
BibTeX
@inproceedings{liu2021deep,
  author = {Yongkai Liu and Qi Miao and Chuthaporn Surawech and Haoxin Zheng and Dan Nguyen and Guang Yang and Steven S. Raman and Kyunghyun Sung},
  title = {{Deep Learning Enables Prostate MRI Segmentation: A Large Cohort Evaluation With Inter-Rater Variability Analysis}},
  booktitle = {Frontiers in Oncology},
  year = {2021},
  doi = {10.3389/fonc.2021.801876},
}
Figure