[论文解读] The state-of-the-art 3D anisotropic intracranial hemorrhage segmentation on non-contrast head CT: The INSTANCE challenge

Xiangyü Li, Gongning Luo+32 more|arXiv (Cornell University)|Jan 9, 2023

Intracerebral and Subarachnoid Hemorrhage Research被引用 19

一句话总结

INSTANCE 2022 MICCAI 挑战标准化在各向异性非对比头颅CT上对颅内出血分割的评估，比较了13种方法在四个指标上的表现，并强调了用于 ICH 分割的数据、方法和瓶颈。

ABSTRACT

Automatic intracranial hemorrhage segmentation in 3D non-contrast head CT (NCCT) scans is significant in clinical practice. Existing hemorrhage segmentation methods usually ignores the anisotropic nature of the NCCT, and are evaluated on different in-house datasets with distinct metrics, making it highly challenging to improve segmentation performance and perform objective comparisons among different methods. The INSTANCE 2022 was a grand challenge held in conjunction with the 2022 International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI). It is intended to resolve the above-mentioned problems and promote the development of both intracranial hemorrhage segmentation and anisotropic data processing. The INSTANCE released a training set of 100 cases with ground-truth and a validation set with 30 cases without ground-truth labels that were available to the participants. A held-out testing set with 70 cases is utilized for the final evaluation and ranking. The methods from different participants are ranked based on four metrics, including Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), Relative Volume Difference (RVD) and Normalized Surface Dice (NSD). A total of 13 teams submitted distinct solutions to resolve the challenges, making several baseline models, pre-processing strategies and anisotropic data processing techniques available to future researchers. The winner method achieved an average DSC of 0.6925, demonstrating a significant growth over our proposed baseline method. To the best of our knowledge, the proposed INSTANCE challenge releases the first intracranial hemorrhage segmentation benchmark, and is also the first challenge that intended to resolve the anisotropic problem in 3D medical image segmentation, which provides new alternatives in these research fields.

研究动机与目标

• 促进在各向异性 3D NCCT 数据上对 ICH 分割方法的公平、客观比较。
• 提供标准化的数据集和颅内出血分割的评估协议。
• 评估各向异性及出血亚型对分割性能的影响。
• 识别瓶颈并指导未来在 ICH 分割和各向异性数据处理方面的改进。

提出的方法

• 整理了包含 refined radiologist 注释的 200 个 NCCT 体积数据集，分为训练/验证/测试集（100/30/70）。
• 定义了四个评估指标（DSC、HD、RVD、NSD）以及跨参与团队的先聚合后排序的评估框架。
• 在 INSTANCE 数据上重新训练了基线 SLEX-NET 模型，以建立参考性能。
• 参与者主要使用基于 U-Net 的架构（特别是 nnU-Net），包含 2D/3D 变体、数据增强、集成以及多样的损失函数。
• 提供基线实现和与 Grand Challenge 兼容的管线，以实现公平比较。

实验结果

研究问题

• RQ1 最先进的 3D 分割方法如何处理用于颅内出血的各向异性 NCCT 体积？
• RQ2 在 INSTANCE 数据上，不同架构（2D/3D/混合）在标准化指标下的基准性能如何？
• RQ3 血肿体积大小和出血亚型如何影响分割精度？
• RQ4 当前 ICH 分割的瓶颈是什么，以及如何改进各向异性处理？

主要发现

团队	DSC(%) 	NSD(%) 	RVD	HD(mm)
T1	79.12 b1 23.00	50.26 b1 19.91	0.21 b1 0.20	29.02 b1 26.34
T2	78.21 b1 18.45	55.28 b1 12.67	0.20 b1 0.18	32.30 b1 30.04
T3	71.60 b1 30.10	50.60 b1 21.30	0.29 b1 0.30	inf
T4	73.55 b1 26.74	51.57 b1 18.10	0.24 b1 0.24	27.16 b1 32.41
T5	73.39 b1 27.38	51.93 b1 18.99	0.25 b1 0.27	inf
T6	79.53 b1 17.18	56.81 b1 12.47	0.20 b1 0.18	21.56 b1 25.02
T7	71.12 b1 29.38	50.19 b1 20.56	0.27 b1 0.30	inf
T8	72.34 b1 28.52	48.93 b1 19.57	0.58 b1 1.65	35.37 b1 29.53
T9	69.96 b1 30.26	48.75 b1 19.66	0.26 b1 0.27	inf
T10	69.28 b1 28.39	46.34 b1 19.54	0.36 b1 0.44	36.23 b1 2.01
T11	52.87 b1 29.66	27.36 b1 14.38	2.16 b1 4.86	149.77 b1 44.52
T12	64.76 b1 31.42	40.26 b1 19.93	0.52 b1 0.76	57.13 b1 22.53
T13	67.16 b1 33.19	45.58 b1 22.35	0.27 b1 0.29	38.88 b1 39.56

• 获胜者在测试数据集上平均 DSC 为 0.6925，显著高于基线。
• 在 13 支队伍中，平均 DSC 范围为 40.22%~72.06%，NSD 为 25.11%~53.59%，RVD 为 0.21~1.55，HD 为 21.56 mm~149.77 mm（有大量无穷大表示检测失败）。
• 基线 SLEX-Net 在测试阶段达到 52.83% DSC 和 0.725 NSD，凸显整体挑战难度。
• 小型出血（低血肿体积）在各方法上显著更难分割，如体积-DSC 分析所示。
• 蛛网膜下腔出血（SAH）在亚型中持续表现最差，指示了一个关键的改进瓶颈。