A structured multi-scale dataset with prostate MRI for AI/ML research

项目总结 磁共振成像(MRI)可以提供前列腺的详细解剖和功能信息， 但放射科医生目前报告的是有限的特征，通常是主观和定性的。近期 人工智能和机器学习(AI/ML)的发展已经证明，AI/ML模型可以 通过学习重要的知识来补充和克服当前定性MRI解释的障碍 从数据中预测具有临床意义的前列腺癌的分层特征和微妙模式。 AI/ML的一个挑战是提供足够数量的有效注释(例如，前列腺癌的轮廓 损害，格里森为每个损害评分)，以确保“地面真相”标签是不偏不倚和生物学的 切合实际。目前，这些地面真实标记通常是从组织病理学证实的 检查结果，可以来自活组织检查或手术标本。然而，这两个组织病理学发现 经常是不和谐的。特别是，已知基于活检的组织病理学结果是有偏见的和/或不确定的 由于(1)病理学家之间的解释差异，(2)具有交界性级别的病变，以及(3)活检 抽样误差。这种差异直接影响AI/ML的训练、验证和推广。 到目前为止，癌症成像档案中存在公开可用的前列腺癌MRI数据集，但这些数据集利用 活检-确认组织病理学为基本事实标签。有必要解决潜在的不确定性和 前列腺癌MRI数据集的数据标记中的偏差。以正在进行的NIH R01项目(R01-CA248506)为基础 正在开发新的定量MRI和AI/ML方法来预测具有临床意义的前列腺癌 在外科病理学方面，本项目的目标是改善前列腺MRI的AI/ML准备情况 通过将来自活组织检查和手术的临床、放射和病理数据的多尺度信息链接起来的数据 在同一群人中。我们的研究团队将通过集成前列腺MRI来构建一个支持人工智能的多尺度数据集 在同一队列中进行不同的组织病理学分析。这将允许直接比较和验证 当使用不同的基础事实标签时，不同的AI/ML模型，提供了潜在的方法来组合其他 公开可用的AI/ML数据集。调查小组增加了核磁共振-超声融合的专家 活组织检查和生物医学信息学，以开发准备用于培训和验证的多尺度数据集 AI/ML算法。拟议工作的成功完成将产生：(1)经同意的唯一数据集 接受了前列腺MRI以及活检和前列腺切除术的受试者；以及(2)结构化临床， 放射学和病理学结果以标准化的方式与明确定义的数据字典共享。这 扩充的人口和工具包将使图像到的AI/ML模型得到进一步完善和改进 主动监测前列腺癌患者的组织病理学相关性和时间监测。

PROJECT SUMMARY Magnetic resonance imaging (MRI) can provide detailed anatomical and functional information of the prostate, but radiologists presently report on limited characteristics, often in a subjective and qualitative manner. Recent developments in artificial intelligence and machine learning (AI/ML) have demonstrated that AI/ML models can complement and overcome the obstacles of current qualitative MRI interpretation by learning important hierarchical features and subtle patterns that are predictive of clinically significant prostate cancer from the data. A challenge in AI/ML is providing an adequate number of validated annotations (e.g., contours of prostate cancer lesions, Gleason scores for each lesion) to ensure that "ground truth" labels are unbiased and biologically relevant. Presently, these ground truth labels are commonly obtained from histopathologically-confirmed findings, which can be from either biopsy or surgical specimens. However, these two histopathological findings are often discordant. Particularly, biopsy-based histopathology results are known to be biased and/or uncertain due to (1) interpretation variability among pathologists, (2) lesions with borderline grades, and (3) biopsy sampling error. This discrepancy directly impacts the training, validation, and generalization of AI/ML. To date, publicly available prostate MRI datasets exist in The Cancer Imaging Archive, but these datasets utilize biopsy-confirmed histopathology as ground truth labels. There is a need to address the potential uncertainty and bias in data labeling of the prostate MRI datasets. Building upon an active NIH R01 project (R01-CA248506) that is developing novel quantitative MRI and AI/ML methods to predict clinically significant prostate cancer with respect to surgical pathology, the objective of this project is to improve the AI/ML-readiness of the prostate MRI data by linking multiscale information across clinical, radiologic, and pathologic data from biopsy and surgery within the same cohort. Our research team will build an AI-ready multiscale dataset by integrating prostate MRI with different histopathology analyses within the same cohort. This will allow direct comparison and validation of different AI/ML models when different ground truth labels are used, providing potential ways to combine other publicly available AI/ML datasets. The investigative team was augmented with experts in MRI-ultrasound fusion biopsy and biomedical informatics to develop a multiscale dataset that is ready for training and validation of AI/ML algorithms. Successful completion of the proposed work will result in: (1) a unique dataset of consented subjects who underwent prostate MRI and both biopsies and prostatectomy; and (2) structured clinical, radiologic, and pathologic findings shared in a standardized manner with a clearly defined data dictionary. This augmented population and toolkit will enable further refinement and improvements in AI/ML models for image to histopathology correlation and temporal monitoring of prostate cancer in patients on active surveillance.