中国科学院分子影像重点实验室知识库

胃癌严重威胁了国民生命安全，为我国医疗系统带来了极大的诊疗负担。治疗前对胃癌分期的精确诊断对临床医师制定治疗方案具有重要的参考价值。计算机断层扫描（Computed Tomography，CT）是术前病情评估中最常用的检查方法，它可以提供一种非侵入性的手段使得医生可以观察到胃部的全貌。然而单纯通过医生肉眼主观判断，很难实现胃癌的精准分期诊断。

影像组学（Radiomics）作为一种新兴的人工智能辅助诊断手段，有潜力准确地描述胃癌特征，并对临床指标进行预测。然而在相关研究中，选择使用三维（3D）感兴趣区域（Regions of Interest，ROI）标注，还是更加节省人力算力的二维（2D） ROI标注，长期以来存在争议。针对这一挑战性问题，本文开展了基于2D-3D多维CT影像组学特征的胃癌分期预测方法研究，探索最优的ROI标注方式和特征提取方式，并做出以下工作：

1. 针对2D和3D影像组学特征哪一个更适合胃癌影像组学临床指标预测的问题，本文通过三个基于胃癌的临床感兴趣事件作为预测指标，通过构建六个模型，研究对比了3D影像组学特征和2D影像组学特征的预测性能。实验结果显示2D模型的效果均优于3D模型，表明了3D ROI的标注方式和3D影像组学特征的计算都不能给模型预测带来收益。
2. 针对上述对比研究可能存在的可复现性问题，本文通过一项模型稳定性研究，建立了45000个模型，探索了在不同重采样间隔、不同数据集划分的情况下，2D和3D模型的性能。结果显示2D模型的效果仍然普遍优于3D模型，进一步拓展和加强了对比研究的结论。
3. 本文进一步结合了2D和3D影像组学特征建立了综合模型，探索能否结合2D和3D特征的优点做出更加准确的预测。结果显示相比于2D模型，综合模型没有增益，从侧面进一步证明了对比研究的结论。

通过本文研究，我们发现3D勾画和3D影像组学特征相比于2D并不能带来收益，2D影像组学模型能够更好地预测胃癌临床指标。这一发现可以在胃癌相关的影像组学分析中为研究者提供特征和算法选择方面的依据。通过使用更加省时省力的2D ROI标注，研究者在保证模型预测性能的同时，可以大大加快相关研究的进展。本文结论有希望促进胃癌影像组学相关研究的流程标准化，减少胃癌影像组学研究壁垒。

以上研究成果以第一作者发表在医学影像领域主流SCI期刊 IEEE Journal of Biomedical and Health Informatics（SCI IF: 5.223）。

Gastric cancer poses a serious threat to national life and brings a huge burden of diagnosis and treatment to our country. Precise diagnosis of gastric cancer staging before treatment has important reference value and influence for clinicians to formulate treatment plan. Computed Tomography (CT), the most commonly used method of preoperative assessment, provides a non-invasive means by which radiologists can observe the full picture of the stomach. However, it is difficult to achieve accurate staging of gastric cancer based solely on the subjective judgment of radiologists.

Radiomics, an emerging artificial intelligence-assisted diagnosis method, has the potential to accurately describe the characteristics of gastric cancer and predict clinical indicators. However, in related studies, the choice to use three-dimensional (3D) Regions of Interest (ROI) labeling, or the more labor-saving two-dimensional (2D) ROI labeling, has long been controversial. In response to this challenging problem, this paper has carried out studies on gastric cancer staging prediction methods based on 2D-3D multi-dimensional CT radiomic features, explored the optimal ROI labeling method and feature extraction method, and made the following work:

1. Aiming at the question of which of the 2D and 3D radiomic features are more suitable for the prediction of gastric cancer staging indicators, three tasks of clinical interest for gastric cancer were conducted (they are, whether the patient has a pathological T4 stage, whether the patient has lymph node metastasis, and whether the patient has vascular invasion), and the predictive performance of 3D radiomic features and 2D radiomic features were compared through six models. The experimental results showed that the effects of 2D models were better than those of 3D models, implying that the annotation of 3D ROI and the calculation of 3D radiomic features could not bring benefits to model performance.

2. In view of the possible reproducibility problem of the main experiment, this project established 45,000 models through an auxiliary experiment, and explored the performance of 2D and 3D models under different resampling intervals and different dataset divisions. The results showed that the effects of the 2D models were still generally better than those of the 3D models, which further expanded and strengthened the conclusion of the main experiment.

3. This project further combined 2D and 3D radiomic features to establish a hybrid model to explore whether the advantages of 2D and 3D features could be combined to make more accurate predictions. The results showed that compared with the 2D model, the hybrid model had no bonus, which further proved the conclusion of the main experiment from the side.

Through this project, we found that 3D ROI annotation and 3D radiomic features could not bring benefits compared to 2D, and the 2D radiomic model could better predict the clinical tasks of gastric cancer. This finding can provide a basis for the researchers to select features and algorithms in the radiomic analysis related to gastric cancer. By using a more time-saving and labor-saving 2D ROI annotation, researchers can significantly speed up the research while maintaining the model's predictive performance. This discovery has the potential to promote the standardization of radiomic research procedures related to gastric cancer and reduce the barriers.

The above research was published in the IEEE Journal of Biomedical and Health Informatics (SCI IF: 5.223) as the first author.