Source-linked AI summary
Deep Learning to Segment Pelvic Bones: Large-scale CT Datasets and Baseline Models
Pengbo Liu, Hu Han, Yuanqi Du, Heqin Zhu, Yinhao Li, Feng Gu, Honghu Xiao, Jun Li, Chunpeng Zhao, Li Xiao, Xinbao Wu, S. Kevin Zhou
TL;DR
Pelvic CT bone segmentation remains limited by appearance variations and the lack of large annotated datasets. The paper curates a multi-domain dataset, trains a multi-class 3D segmentation network, and adds SDF post-processing, achieving an average Dice of 0.987 for metal-free volumes.
Problem
Existing pelvic CT segmentation methods have limited accuracy under varied appearances, while large annotated datasets for exploring deep learning remain lacking.
Method
The paper curates a multi-domain pelvic CT dataset, trains a deep multi-class network for four pelvic bone structures, and applies signed-distance-function post-processing.
Results
Average Dice reaches 0.987 for metal-free volumes, while the 3D U-Net cascade achieves an average Hausdorff distance of 5.50 voxels.
Takeaways & Limitations
The open dataset and baseline system provide a foundation for further pelvic bone segmentation research and planned clinical evaluation.
Takeaways & Limitations
The study reports ethics approval from a clinical hospital, but no further author-supported limitation is stated in the supplied passages.
Abstract
from arXiv · showhide
Purpose: Pelvic bone segmentation in CT has always been an essential step in clinical diagnosis and surgery planning of pelvic bone diseases. Existing methods for pelvic bone segmentation are either hand-crafted or semi-automatic and achieve limited accuracy when dealing with image appearance variations due to the multi-site domain shift, the presence of contrasted vessels, coprolith and chyme, bone fractures, low dose, metal artifacts, etc. Due to the lack of a large-scale pelvic CT dataset with annotations, deep learning methods are not fully explored. Methods: In this paper, we aim to bridge the data gap by curating a large pelvic CT dataset pooled from multiple sources and different manufacturers, including 1, 184 CT volumes and over 320, 000 slices with different resolutions and a variety of the above-mentioned appearance variations. Then we propose for the first time, to the best of our knowledge, to learn a deep multi-class network for segmenting lumbar spine, sacrum, left hip, and right hip, from multiple-domain images simultaneously to obtain more effective and robust feature representations. Finally, we introduce a post-processing tool based on the signed distance function (SDF) to eliminate false predictions while retaining correctly predicted bone fragments. Results: Extensive experiments on our dataset demonstrate the effectiveness of our automatic method, achieving an average Dice of 0.987 for a metal-free volume. SDF post-processor yields a decrease of 10.5% in hausdorff distance by maintaining important bone fragments in post-processing phase. Conclusion: We believe this large-scale dataset will promote the development of the whole community and plan to open source the images, annotations, codes, and trained baseline models at https://github.com/ICT-MIRACLE-lab/CTPelvic1K.
1 Introduction
Pelvic CT segmentation remains difficult because existing methods have limited accuracy under varied imaging conditions, while large annotated datasets for deep learning are lacking. This paper addresses the gap with a multi-domain pelvic CT dataset and deep multi-class segmentation system.
- 1 Introduction: Existing pelvic bone segmentation methods have limited accuracy under density variation, weak fracture edges, and diverse CT appearance conditions.Challenges include domain shift, contrasted vessels, coprolith and chyme, low dose, and metal artifacts.
- 1 Introduction: The paper curates and releases a pelvic CT dataset containing 1,184 volumes and over 320K slices with diverse appearance variations.The dataset includes 75 CTs with metal artifacts and expert multi-bone labels.
- 1 Introduction: The authors learn a multi-class network jointly segmenting lumbar spine, sacrum, left hip, and right hip from multi-domain images.The goal is to obtain more effective and robust representations for pelvic bone segmentation.
- 1 Introduction: The proposed work presents a fully automatic pelvic segmentation pipeline intended to provide accurate, efficient, and robust analysis.The authors identify it as the first real attempt of this type to their knowledge.
2 Our Dataset
The dataset combines pelvic CT scans from seven sources and uses iterative deep learning with expert correction to scale annotation. The resulting annotations cover most metal-free scans, while some metal-affected scans remain reserved for unsupervised learning.
- 2 Our Dataset: The dataset pools pelvic CT images from seven sources with different sites and clinical characteristics to represent practical appearance variations.Images are curated from two clinical sources and five existing CT datasets, with unrelated areas outside the pelvis removed.
- 2 Our Dataset: Annotation by Iterative Deep Learning trains a network on precise initial labels, predicts new cases, and iteratively incorporates human corrections.The workflow is designed to reduce the time required for large-scale subjective annotation.
- 2 Our Dataset: The annotation pipeline begins with 40 precisely labeled cases, processes batches of 100 new cases, and uses junior annotators, coordinators, and senior experts for refinement.Hard cases receive more precise senior-expert annotation before iterative processing continues.
- 2 Our Dataset: Annotations are available for 1,109 metal-free CTs and 14 metal-affected CTs, while 61 additional metal-affected CTs remain unannotated for planned unsupervised learning.The annotation strategy progressively reduces the manual effort needed from experts.
3 Segmentation Methodology
The method combines a 3D U-Net cascade for multi-class pelvic bone segmentation with SDF-based post-processing that removes false predictions while preserving fracture fragments.
- The pipeline sends a 3D CT volume through a replaceable segmentation module and then an SDF post-processor to produce the final multi-bone result.
- 3.1 Segmentation Module: A 3D U-Net cascade segments the lumbar spine, sacrum, left hip, and right hip from 3D CT volumes.The first U-Net processes downsampled images for broader context, while the second refines predictions at full resolution.
- Table 2 evaluates model variants on the six metal-free sub-datasets using Dice coefficient and Hausdorff distance, with average and whole-bone measurements.
- 3.2 SDF Post Processor: SDF filtering adds a distance constraint to connected-region size filtering, helping retain fractured bone fragments that conventional outlier removal may discard.The distance map is computed from the maximum connected region of the predicted anatomy.
4 Experiments
Experiments evaluate implementation details, model performance, cross-subdataset generalization, and SDF post-processing on the pelvic CT dataset.
- Performance of baseline models: 0.987 average DC and 5.50-voxel HD were achieved by ΦALL(3D cascade) on the full testing set.The authors report that the 3D U-Net cascade outperformed the evaluated 2D/3D alternatives for learning pelvic anatomy features.
- Generalization across sub-datasets: Models trained on all sub-datasets generally outperformed models trained on individual sub-datasets across evaluation domains.The comparison used average DC and HD across different datasets and models.
- Generalization across sub-datasets: The experiments include qualitative visualizations across six datasets, with colors denoting the sacrum, left hip, right hip, and lumbar spine.
- Generalization across sub-datasets: Single-subdataset models were difficult to generalize consistently, whereas COLONOG was an exception because it contained the largest amount of data from varied sources.
- Generalization across sub-datasets: Aggregating five sub-datasets achieved a result on unseen COLONOG data comparable to ΦALL while using one third of the data.The leave-one-out experiment supports multi-source training as a way to improve generalization across domains.
- SDF post-processor: 80.7% and 15.1% decreases in HD were obtained with SDF post-processing compared with no post-processor and MCR post-processing, respectively.SDF retained large fragments near anatomical structures that MCR removed.
5 Conclusion
The paper presents an open, multi-domain pelvic CT dataset and a deep multi-class segmentation system with SDF filtering. The authors position these resources as a foundation for future clinical evaluation and further method development.
- 5 Conclusion: The dataset contains 1,184 CT volumes and over 320K slices spanning varied appearance conditions, and is open sourced with the segmentation system.The authors describe the collection as pooled from multiple domains and intended to benefit pelvic surgery and diagnosis research.
- 5 Conclusion: The system segments lumbar spine, sacrum, left hip, and right hip using a multi-class network trained on multiple-domain images.The design targets more effective and robust feature representations across the collected domains.
- 5 Conclusion: SDF filtering further improves the robustness of the pelvic segmentation system.The conclusion identifies SDF filtering as an additional robustness component beyond the multi-class network.
- 5 Conclusion: The authors plan to test the system in real clinical practice and explore metal-affected CT modules and domain-independent segmentation algorithms.These directions are presented as future work based on the released dataset and baseline system.
Declarations
The declarations report funding, conflicts of interest, data and code availability, ethics approval, and the absence of an informed-consent requirement.
- Declarations: The research reports support from the Youth Innovation Promotion Association CAS through grant 2018135.
- Declarations: The authors declare no relevant financial or non-financial conflicts of interest.
- Declarations: Data and code availability are directed to the referenced URL.
- Declarations: The study received approval from the clinical hospital ethics committee, and informed consent was not applicable.