Deep Learning to Segment Pelvic Bones: Large-scale CT Datasets and Baseline Models (2012.08721v2)

Abstract: 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.

• The paper introduces a novel large-scale CT dataset of 1,184 volumes that enhances deep learning for pelvic bone segmentation.
• It employs a deep multi-class 3D U-Net cascade with post-processing via a signed distance function to reduce the Hausdorff distance by 15.1%.
• Results indicate an average Dice coefficient of 0.987, demonstrating high precision and robust generalization across varied imaging conditions.

Deep Learning to Segment Pelvic Bones: Large-scale CT Datasets and Baseline Models

The paper "Deep Learning to Segment Pelvic Bones: Large-scale CT Datasets and Baseline Models" addresses the critical task of advancing automatic pelvic bone segmentation in CT imaging. Recognizing the limitations of existing handcrafted and semi-automatic models, particularly in terms of accuracy under varied imaging conditions, the authors propose a novel approach leveraging deep learning. The paper is distinguished by its attempt to overcome the scarcity of large annotated datasets, a common bottleneck in improving segmentation accuracy.

Contribution and Methodology

One of the paper's primary contributions is the assembly of a large, multi-source pelvic CT dataset, comprising 1,184 volumes (including over 320,000 slices). This dataset includes varied image conditions such as domain shifts from different acquisition sites, the presence of contrasted vessels, artifacts due to metal implants, and various imaging resolutions and fields-of-view. The creation of this dataset is intended to serve as a substantial resource for future research in the medical imaging community.

The authors propose a deep multi-class network capable of segmenting the lumbar spine, sacrum, left hip, and right hip bones. The approach stands out in its dual focus: leveraging a comprehensive dataset for training and employing a robust post-processing method based on the signed distance function (SDF) to refine segmentation outputs. By incorporating SDF, the method aims to improve upon traditional post-processing, specifically reducing the Hausdorff distance by 15.1%, as reported. This reflects more accurate boundary detection against the inherent challenges of noise and artifacts found in CT images.

Results and Discussion

In terms of performance, the proposed model demonstrates exemplary results with an average Dice coefficient of 0.987 on metal-free volumes. This indicates a high level of accuracy in delineating pelvic structures compared to the prevailing methodologies. The results substantiate the hypothesis that large scale and diverse training datasets can significantly improve model robustness and generalization capabilities.

The use of a deep learning model, particularly the 3D U-Net cascade, is validated through extensive experimentation on various sub-datasets within the larger dataset. The experiments show how training on a diverse dataset yields a model with superior generalization across different imaging domains. The paper's methodological rigor, demonstrated through its thorough cross-validation, fortifies its claims about the robustness of its proposed solutions.

Implications and Future Work

From a practical perspective, the development and open-sourcing of this large-scale dataset, alongside the baseline models, potentially catalyze advancements across the medical imaging domain. The dataset's variability prepares models to cope with real-world clinical scenarios, tackling domain adaptation issues often encountered in medical image segmentation tasks.

Theoretically, this research contributes to the exploration of deep learning's capacity in medical imaging, particularly in consistently achieving accuracy across disparate imaging conditions. By emphasizing the impact of dataset diversity on learning efficiency, the paper aligns with broader findings in machine learning regarding the importance of comprehensive training datasets.

Looking to the future, the authors express the intention to extend their framework, addressing unsupervised learning approaches for metal-affected images and developing domain-adaptive algorithms. This progression could see their methodology accommodate even greater complexity within clinical environments, marking a step towards fully integrated AI systems in medical diagnostics and treatment planning.

The overarching impact of this paper is in setting a benchmark for future research while offering tangible resources to further innovations in CT image segmentation. As the medical field increasingly relies on AI and machine learning for diagnostics, the systems and datasets spearheaded by studies such as this will underpin developments aligning with accuracy, efficiency, and generalizable applicability in clinical settings.