PELE scores: Pelvic X-ray Landmark Detection by Pelvis Extraction and Enhancement

Zhen Huang, Han Li, Shitong Shao, Heqin Zhu, Huijie Hu, Zhiwei Cheng, Jianji Wang, S. Kevin Zhou

TL;DR

This paper introduces PELE, a novel two-stage method for pelvic X-ray (PXR) landmark detection. PELE first extracts the pelvis bone from PXRs using 3D CT priors and then enhances the extracted bone image. This approach significantly improves landmark detection accuracy by overcoming the challenge of soft tissue superposition in PXRs, achieving state-of-the-art results.

What Problem Does This Paper Address?

Pelvic X-rays (PXRs) are a common and cost-effective tool for diagnosing pelvic diseases. However, a major challenge in analyzing PXRs is the "pelvis-tissue superposition" problem. This occurs because the 2D nature of X-rays causes soft tissues like intestines, bladder, and urethra to overlap with the pelvic bones. This overlap can obscure important anatomical features and confuse automated landmark detection systems, leading to inaccurate diagnoses and treatments. Existing landmark detection methods primarily focus on network architecture and feature learning, neglecting the crucial step of improving image visibility for the target anatomy.

Methods

The proposed PELE (PELvis Extraction) module tackles this problem in two stages:

  1. Pelvis Extraction: This stage leverages 3D CT scans to guide the extraction of the pelvic bone from 2D PXRs.
  2. Image Decomposition: The authors first use 3D CT volumes (both raw and with isolated pelvis) to generate 2D Digitally Reconstructed Radiographs (DRRs). A deep neural network (FDE) is trained to decompose a DRR image into a bone-only image and a tissue-only image, guided by a bone mask segmentation task.
  3. Domain Adaptation: Since DRRs and PXRs have a domain gap, a CycleGAN-based domain adaptation network (FDA) is employed to bridge this gap, allowing the trained decomposition model to work on real PXRs. The final extracted pelvis bone image is denoted as $X_{bone}$.
  4. Pelvis Enhancement: To further improve the quality of the extracted pelvis, an enhancement module ($F_{enh}$) is introduced.
  5. This module smooths the contours of the extracted pelvis and then combines it with the original PXR to reintroduce finer details. Tone mapping techniques are applied to enhance visibility in darker areas. The output is an enhanced pelvis image, $X_{boneEnh}$.

Finally, the enhanced pelvis image ($X_{boneEnh}$) is used as input for a standard landmark detection network, which is trained using cross-entropy loss.

Datasets

The study utilizes several datasets for training and evaluation:

  • CT Datasets: The public CTPelvic1K dataset (1184 volumes) was used to generate DRR images for training the domain adaptation and decomposition networks.
  • PXR Datasets:
    • CGMH-PelvisSeg dataset (400 high-resolution PXRs).
  • A dataset from a previous study [5] (150 images).
  • An in-house private dataset (300 images).
  • A total of 850 PXRs were used for evaluation.

Key Insights and Ideas

  • The "Pelvis-Tissue Superposition" is a critical bottleneck: The paper strongly argues that simply improving landmark detection algorithms isn't enough; addressing the inherent visibility issues in PXRs is paramount.
  • Leveraging 3D CT priors for 2D PXR analysis: The innovative use of 3D CT data to guide the segmentation of 2D PXRs is a key contribution, effectively transferring anatomical knowledge across modalities.
  • Enhancement is crucial: The enhancement module ($F_{enh}$) is shown to be indispensable, further improving the quality of the extracted pelvis and leading to better landmark detection.
  • PELE benefits diverse landmark detectors: The method is shown to improve the performance of different baseline landmark detection networks (U-Net and GU2Net), demonstrating its generalizability.
  • Significant gains on small datasets: PELE shows a particularly large improvement when trained on limited data, suggesting its potential for scenarios where large annotated datasets are scarce.

Key Results

  • Improved Landmark Detection Accuracy: PELE significantly reduces Mean Radial Error (MRE) and increases Successful Detection Rate (SDR) for pelvic landmark detection. For example, with 425 PXRs, PELE achieved an MRE of 1.83px and 94.41% SDR within a 9px distance.
  • State-of-the-Art Performance: The proposed PELE module achieves state-of-the-art performance on benchmark metrics for pelvic landmark detection.
  • Superior Bone Extraction: Quantitative metrics (Dice, HD95, SSIM, LPIPS) show that PELE's pelvis extraction outperforms baseline methods like U-Net and CycleGAN. For instance, PELE achieved a Dice score of 0.945 and an SSIM of 0.923.
  • Qualitative Improvements: Visualizations clearly demonstrate that PELE effectively isolates the pelvis bone and enhances its details, leading to more accurate landmark predictions.

Limitations and Future Work

  • Limitations: The paper acknowledges that even with PELE, the standard deviation of landmark detection can be high (9.32px), and experts still find some PXRs challenging to mark accurately due to extreme superstition. The performance of GU2Net without PELE was noted to degenerate due to not using multi-organ datasets during training.
  • Future Work: The authors plan to investigate PELE's potential for other downstream clinical tasks beyond landmark detection, such as surgical planning and pathological diagnosis. They also aim to explore additional high-quality datasets and design improved networks to further enhance their results.

Summary updated

Abstract

The pelvis, the lower part of the trunk, supports and balances the trunk. Landmark detection from a pelvic X-ray (PXR) facilitates downstream analysis and computer-assisted diagnosis and treatment of pelvic diseases. Although PXRs have the advantages of low radiation and reduced cost compared to computed tomography (CT) images, their 2D pelvis-tissue superposition of 3D structures confuses clinical decision-making. In this paper, we propose a PELvis Extraction (PELE) module that utilizes 3D prior anatomical knowledge in CT to guide and well isolate the pelvis from PXRs, thereby eliminating the influence of soft tissue. We conduct an extensive evaluation based on two public datasets and one private dataset, totaling 850 PXRs. The experimental results show that the proposed PELE module significantly improves the accuracy of PXRs landmark detection and achieves state-of-the-art performances in several benchmark metrics, thus better serving downstream tasks.
BibTeX
@misc{huang2023pele,
  author = {Zhen Huang and Han Li and Shitong Shao and Heqin Zhu and Huijie Hu and Zhiwei Cheng and Jianji Wang and S. Kevin Zhou},
  title = {{PELE scores: Pelvic X-ray Landmark Detection by Pelvis Extraction and Enhancement}},
  year = {2023},
  eprint = {2305.04294},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.04294},
}

Loading related content...

Figure