TL;DR
This paper introduces PELE, a novel two-stage module for pelvic X-ray (PXR) landmark detection that explicitly addresses the challenge of soft tissue superposition. By first extracting and enhancing the pelvis bone from PXRs using 3D CT priors and domain adaptation, PELE significantly improves landmark detection accuracy, achieving state-of-the-art results.
What Problem Does This Paper Address?
Pelvic X-rays (PXRs) are crucial for diagnosing and treating pelvic conditions due to their low radiation and cost. However, a major hurdle in accurately analyzing PXRs is the "pelvis-tissue superposition" problem. This occurs because the 2D X-ray image captures the pelvis overlaid with soft tissues like intestines and bladder. This superposition can obscure critical anatomical landmarks, leading to errors in automated landmark detection systems. While existing deep learning methods try to implicitly handle this by designing more complex networks, this paper argues for an explicit approach: first isolating the pelvis bone itself before attempting landmark detection.
Methods
The core of the proposed solution is the PELvis Extraction (PELE) module, which operates in two main stages:
- Image Decomposition: This stage aims to separate the pelvis bone from the soft tissues in a PXR. Since paired 2D PXR and 3D CT data with isolated pelvises are scarce, the authors leverage 3D CT volumes. They generate 2D Digitally Reconstructed Radiographs (DRRs) from 3D CT data (both the full volume and the isolated pelvis). A deep neural network, FDE, is trained to decompose a DRR of the full pelvis into a pelvis-only DRR. To aid this decomposition, they also incorporate a bone mask segmentation task.
- Domain Adaptation (DA): A domain gap exists between DRR images (generated from CT) and actual PXRs. To bridge this gap, the authors employ a CycleGAN-based network, FDA, to learn a mapping from PXR images to DRR images. This allows the decomposition model (trained on DRRs) to be applied to real PXRs.
- Pelvis Enhancement: After extracting the pelvis bone (resulting in
Xbone), an enhancement module, Fenh, is applied. This module smooths transitions, normalizes contours, and uses tone mapping to improve the visibility of less-penetrated areas (like hip and sacrum bones) that might be artifacts or poorly rendered in the initial extraction.
The final enhanced pelvis image (XboneEnh) is then fed into a standard landmark detection network for improved accuracy.
Datasets
The study utilizes a combination of datasets for training and evaluation:
- CT Datasets: The public CTPelvic1K dataset (1184 volumes, >320K slices) is used to generate DRR images for training the decomposition and domain adaptation modules.
- PXR Datasets:
- Dataset 1: Open-source CGMH-PelvisSeg (400 high-resolution PXRs).
- Dataset 2: An open-source dataset from a previous study [8] (150 images).
- Dataset 3: An in-house dataset of 300 retrospectively collected images from cooperative hospitals.
In total, 850 PXRs from these three sources were used for evaluation.
Key Insights and Ideas
- Explicitly Handling Superposition: The paper's core insight is that implicitly handling soft tissue superposition in PXR landmark detection is insufficient. An explicit pelvis extraction step is crucial for robust performance.
- Leveraging 3D CT Priors: The scarcity of paired 2D/3D data is cleverly overcome by using 3D CT volumes to generate synthetic DRR data, which then bridges the gap to real PXRs via domain adaptation.
- Domain Adaptation is Key: The domain gap between synthetic DRRs and real PXRs is a significant challenge. The use of CycleGAN for domain adaptation proves effective in transferring learned pelvis features to real-world images.
- Enhancement Matters: Even after bone extraction, artifacts or poor contrast can persist. The enhancement module (
Fenh) plays a vital role in improving image quality and making critical bone structures more visible. - Generalizability: The PELE module demonstrates effectiveness across different landmark detection baselines (U-Net, GU2Net), suggesting its broad applicability.
Key Results
- Improved Landmark Detection Accuracy: The PELE module significantly reduces the Mean Radial Error (MRE) for landmark detection. For example, with 50% of training data, the MRE dropped from 3.41px (baseline U-Net) to 1.83px (U-Net with PELE).
- State-of-the-Art Performance: The proposed method achieved state-of-the-art results on benchmark metrics, with a notable MRE of 0.71 mm compared to 0.93 mm from a previous study using a much larger dataset.
- Enhanced Image Quality: Subjective evaluations by experts showed that the enhanced pelvis images (
XboneEnh) provided higher diagnostic value than the original PXRs (Xraw) or just the extracted bone (Xbone). - Robustness: The method shows significant improvements even with limited training data (e.g., a >200% MRE improvement with only 12.5% of PXRs).
- Quantitative Performance: Table 1 shows PELE achieving superior Dice (0.945), HD95 (34.723), SSIM (0.923), and LPIPS (0.083) scores for pelvis extraction compared to U-Net and CycleGAN baselines.
Limitations and Future Work
- Inter-Expert Variability: The paper acknowledges that even expert annotators can have difficulty marking landmarks on extremely superimposed PXRs, leading to variations in ground truth and affecting evaluation.
- Computational Cost: While not explicitly detailed as a limitation, the multi-stage approach (decomposition, adaptation, enhancement, detection) implies a higher computational cost than a single-stage detector.
- Future Work: The authors suggest investigating the significance of their method for other clinical tasks beyond landmark detection, implying its potential for broader applications in PXR analysis. They also plan to explore its use in diagnosis and treatment.
Summary updated