Accurate Prostate Segmentation in Large-Scale Magnetic Resonance Imaging Datasets via First-in-First-Out Feature Memory and Multi-Scale Context Modeling

Jingyi Zhu, Xukun Zhang, Xiao Luo, Zhiji Zheng, Kun Zhou, Yanlan Kang, Haiqing Li, Daoying GengJournal of Imaging 2025 · DOI

TL;DR

This paper introduces MCM-UNet, a novel deep learning architecture for accurate prostate segmentation in MRI scans. By incorporating a multi-scale context modeling module and a first-in-first-out feature memory bank, MCM-UNet significantly improves segmentation accuracy, especially in challenging regions like the prostate apex and base, outperforming existing methods.

What Problem Does This Paper Address?

Prostate cancer is a major global health concern, and accurate segmentation of the prostate in MRI scans is crucial for effective diagnosis, treatment planning (like radiation therapy), and prognosis. However, segmenting the prostate from surrounding soft tissues in MRI images is challenging due to several factors:

  • Subtle Boundaries: The distinction between the prostate and adjacent tissues can be indistinct, making it hard for algorithms to precisely define the organ's edges.
  • Blurred Regions: The apex and base of the prostate often exhibit blurriness in MRI, further complicating accurate edge detection and segmentation.
  • Inter-Operator Variability: Manual segmentation by radiologists, while the gold standard, is time-consuming and prone to variations between different annotators.

These challenges necessitate automated methods that can achieve high accuracy and consistency, especially in the difficult-to-segment peripheral regions.

Methods

The core of this work is the MCM-UNet architecture, which builds upon the U-Net framework. It introduces two key novel components:

  1. Multi-Scale Context Modeling (MCM) Module: This module is integrated into the skip connections of the U-Net. It aims to refine pixel representations by extracting contextual semantic information across different scales within an image. By selectively reducing the influence of irrelevant features, it enhances the representation of boundary pixels, leading to more precise segmentation, particularly at the prostate's edges.
  2. First-in-First-Out (FIFO) Feature Update Strategy with Memory Bank: This component addresses the need for robust, dataset-level semantic features. It maintains a memory bank that dynamically stores and updates feature vectors from the training data. When processing a new image, the system compares its image-level features with those in the memory bank. It then fuses the most similar dataset-level features with the current image features to create an updated feature representation. This strategy helps capture long-range semantic dependencies and improves segmentation in challenging regions like the apex and base by leveraging information from the entire dataset.

The network uses a standard encoder-decoder structure with skip connections, enhanced by these novel modules. The loss function is a combination of Dice loss and cross-entropy loss, chosen for its effectiveness in emphasizing overlap and penalizing large discrepancies.

Datasets

The study utilizes two datasets:

  • Private Dataset: A large, newly compiled dataset consisting of 2175 T2W MRI scans from 14 different public hospitals in China. This dataset includes scans from both healthy individuals and prostate cancer patients at various stages. It is highlighted as the largest private T2W prostate dataset to date, contributing to the model's robustness and generalization.
  • Public Dataset: The PROMISE12 dataset, a well-established benchmark for prostate segmentation, containing 50 MRI T2W scans. This dataset was used for external validation to assess the model's performance on unseen data from a different source.

Key Insights and Ideas

  • Context is King (Especially Multi-Scale): The MCM module's success demonstrates that simply looking at local pixel information isn't enough for precise segmentation. Capturing broader contextual relationships within an image, especially at multiple scales, is crucial for understanding organ boundaries.
  • Leveraging the Collective Wisdom (Dataset-Level Features): The FIFO memory bank acts like a collective memory of the dataset. By dynamically updating and referencing these "learned experiences," the model can better handle variations and ambiguities present across many scans, improving performance on difficult cases.
  • Apex and Base are the Toughest: The paper consistently emphasizes the challenges in segmenting the apex and base of the prostate. The proposed method's success in these specific regions is a strong indicator of its effectiveness in handling real-world clinical complexities.
  • Large, Diverse Datasets Matter: The creation and use of a large, multi-center private dataset are significant. This approach helps ensure the model's findings are not just specific to one type of scanner or population, leading to more reliable and generalizable results.

Key Results

The MCM-UNet demonstrated superior performance compared to several state-of-the-art methods on both the private and PROMISE12 datasets.

  • Quantitative Performance:
    • On the private dataset, MCM-UNet achieved an Average Symmetric Surface Distance (ASSD) of 0.58 voxels and a Dice Similarity Coefficient (DSC) of 91.71%. This represents a significant improvement over the baseline nnU-Net (ASSD: 1.01 voxels, DSC: 88.13%).
  • On the PROMISE12 dataset, it achieved an ASSD of 1.07 voxels and a DSC of 90.47%, also outperforming other methods.
  • Qualitative Findings: Visualizations showed that MCM-UNet successfully segmented the entire prostate, with minimal under- or over-segmentation, particularly in the challenging apex and base regions. The segmentation results closely matched those of experienced radiologists.

Limitations and Future Work

  • Focus on T2W Modality: The current study exclusively uses T2-weighted MRI sequences. The authors acknowledge that incorporating multiparametric MRI (e.g., T1-weighted, diffusion-weighted) could further enhance segmentation accuracy, as these modalities provide complementary information.
  • Clinical Integration: For widespread clinical adoption, the model needs to be validated on a more diverse range of clinical datasets, integrated with existing PACS systems, and optimized for real-time inference speed. Developing a user-friendly interface for radiologists is also crucial.
  • Thin-Sliced Data: While the model performed well, the authors note that differences in performance metrics between datasets might be influenced by slice thickness. Future work could involve incorporating more thin-sliced data to further diversify experimental datasets.
  • T1W Contrast: The authors mention that T1W images typically offer poor contrast to prostate tissue, suggesting that channel weighting will be important if T1W images are integrated in future studies.

Summary updated

Abstract

Prostate cancer, a prevalent malignancy affecting males globally, underscores the critical need for precise prostate segmentation in diagnostic imaging. However, accurate delineation via MRI still faces several challenges: (1) The distinction of the prostate from surrounding soft tissues is impeded by subtle boundaries in MRI images. (2) Regions such as the apex and base of the prostate exhibit inherent blurriness, which complicates edge extraction and precise segmentation. The objective of this study was to precisely delineate the borders of the prostate including the apex and base regions. This study introduces a multi-scale context modeling module to enhance boundary pixel representation, thus reducing the impact of irrelevant features on segmentation outcomes. Utilizing a first-in-first-out dynamic adjustment mechanism, the proposed methodology optimizes feature vector selection, thereby enhancing segmentation outcomes for challenging apex and base regions of the prostate. Segmentation of the prostate on 2175 clinically annotated MRI datasets demonstrated that our proposed MCM-UNet outperforms existing methods. The Average Symmetric Surface Distance (ASSD) and Dice similarity coefficient (DSC) for prostate segmentation were 0.58 voxels and 91.71%, respectively. The prostate segmentation results closely matched those manually delineated by experienced radiologists. Consequently, our method significantly enhances the accuracy of prostate segmentation and holds substantial significance in the diagnosis and treatment of prostate cancer.
BibTeX
@article{zhu2025accurate,
  author = {Jingyi Zhu and Xukun Zhang and Xiao Luo and Zhiji Zheng and Kun Zhou and Yanlan Kang and Haiqing Li and Daoying Geng},
  title = {{Accurate Prostate Segmentation in Large-Scale Magnetic Resonance Imaging Datasets via First-in-First-Out Feature Memory and Multi-Scale Context Modeling}},
  journal = {Journal of Imaging},
  year = {2025},
  doi = {10.3390/jimaging11020061},
}
Figure