[论文解读] PETWB-REP: A Dataset of Whole-body PET/CT Scans with Paired Radiology Reports

Yichi Zhang, Le Xue+6 more|ArXiv.org|Feb 20, 2025

Medical Imaging Techniques and Applications被引用 4

一句话总结

Paper 介绍了 SegAnyPET，一种用于 PET 图像的三维可提示分割基底模型，在 PETS-5k 数据集上进行训练，能够从高质量和低质量注释中实现鲁棒学习，并对未见器官和数据集具有强泛化能力。

ABSTRACT

Positron Emission Tomography (PET) is a powerful molecular imaging tool that plays a crucial role in modern medical diagnostics by visualizing radio-tracer distribution to reveal physiological processes. Accurate organ segmentation from PET images is essential for comprehensive multi-systemic analysis of interactions between different organs and pathologies. Existing segmentation methods are limited by insufficient annotation data and varying levels of annotation, resulting in weak generalization ability and difficulty in clinical application. Recent developments in segmentation foundation models have shown superior versatility across diverse segmentation tasks. Despite the efforts of medical adaptations, these works primarily focus on structural medical images with detailed physiological structural information and exhibit limited generalization performance on molecular PET imaging. In this paper, we collect and construct PETS-5k, the largest PET segmentation dataset to date, comprising 5,731 three-dimensional whole-body PET images and encompassing over 1.3M 2D images. Based on the established dataset, we develop SegAnyPET, a modality-specific 3D foundation model for universal promptable segmentation from PET images. To issue the challenge of discrepant annotation quality, we adopt a cross prompting confident learning (CPCL) strategy with an uncertainty-guided self-rectification process to robustly learn segmentation from high-quality labeled data and low-quality noisy labeled data for promptable segmentation. Experimental results demonstrate that SegAnyPET can segment seen and unseen target organs using only one or a few prompt points, outperforming state-of-the-art foundation models and task-specific fully supervised models with higher accuracy and strong generalization ability for universal segmentation.

研究动机与目标

• Motivate robust, universal segmentation for PET images with low contrast and weak boundaries.
• Create a large-scale, whole-body PET segmentation dataset (PETS-5k) to enable a PET-specific foundation model.
• Develop a 3D promptable segmentation architecture tailored to PET volumes.
• Address annotation quality variability with a noise-robust training strategy.
• Demonstrate strong generalization to unseen organs and external PET datasets.

提出的方法

• Construct PETS-5k, the largest 3D PET segmentation dataset to date (5,731 PET volumes, >1.3M 2D slices).
• Develop SegAnyPET, a modality-specific 3D segmentation foundation model with image encoder, prompt encoder, and mask decoder.
• Reformulate a 3D architecture to exploit volumetric context for universal segmentation from PET images.
• Adopt cross prompting confident learning (CPCL) to learn from high-quality (HQ) and noisy low-quality (LQ) annotations.
• Use uncertainty-guided self-rectification to refine noisy labels and improve training on LQ data.
• Employ a training objective that blends HQ supervised loss, CPCL consistency loss, and rectified LQ supervised loss.

实验结果

研究问题

• RQ1 Can SegAnyPET achieve accurate universal segmentation on PET images with minimal prompting?
• RQ2 Does a 3D PET-specific foundation model generalize to unseen organs and out-of-distribution datasets?
• RQ3 How does CPCL with uncertainty-guided rectification perform when learning from high- and low-quality annotations?

主要发现

Method	Prompt	Liver	Kidney-L	Kidney-R	Heart	Spleen	Avg
SAM	1 point	26.55	9.38	9.10	14.44	6.30	13.15
MedSAM	1 point	0.25	0.19	1.32	0.27	0.27	0.46
SAM-Med3D	1 point	51.63	21.01	19.17	60.11	25.41	35.46
SAM-Med3D-organ	1 point	80.25	44.70	35.76	74.00	69.23	60.79
SAM-Med3D-turbo	1 point	79.46	66.95	72.81	73.03	68.19	72.09
SegAnyPET	1 point	93.06	89.84	90.61	88.29	90.67	90.49
SAM	3N points	43.85	23.21	22.16	29.09	11.83	26.03
MedSAM	3N points	26.59	28.86	28.98	18.82	32.96	27.24
SAM-Med3D	3 points	62.15	28.21	31.19	61.44	27.07	42.01
SAM-Med3D-organ	3 points	84.82	47.33	48.57	75.85	74.60	66.23
SAM-Med3D-turbo	3 points	84.11	74.05	76.17	75.24	73.34	76.58
SegAnyPET	3 points	93.36	90.25	90.95	88.86	91.10	90.90
SAM	5N points	54.49	47.16	37.42	42.19	18.79	40.01
MedSAM	5N points	36.53	37.53	39.22	24.71	41.30	35.86
SAM-Med3D	5 points	61.05	31.05	31.98	61.88	29.75	43.14
SAM-Med3D-organ	5 points	85.52	49.56	54.40	76.30	75.13	68.18
SAM-Med3D-turbo	5 points	85.56	76.74	78.08	76.16	75.20	78.35
SegAnyPET	5 points	93.42	90.39	91.24	88.95	91.22	91.05

• SegAnyPET substantially outperforms state-of-the-art segmentation foundation models and task-specific models in zero-shot promptable PET segmentation.
• With one point prompt, SegAnyPET achieves average DSC around 90+% on seen organs (Liver, Kidney-L, Kidney-R, Heart, Spleen) in Table 1; 3 points yields comparable scores; 5 points maintains high performance.
• SegAnyPET shows strong generalization to unseen training-invisible organs and to the AutoPET-Organ external dataset (e.g., Table 3 results).
• CPCL with consistency regularization and uncertainty-guided label rectification improves learning from noisy LQ annotations (Table 4 ablations).
• PETS-5k is the largest public 3D PET segmentation dataset to date, enabling the first PET-focused segmentation foundation model with strong performance.
• The approach emphasizes efficient prompts (one or few points) to achieve accurate segmentation, reducing manual effort.