Over the past decade, the medical imaging literature has been revolutionized by radiomics, which allows to extract huge amounts of quantitative data from magnetic resonance images and other imaging modalities for a wide variety of diagnostic and prognostic purposes, particularly in oncology. Radiomics can capture imaging features of tumors that go beyond the conventional visual analysis and well correlate with histologic types, genetic mutations, response to treatment, and patient survival.¹

Despite promising results achieved by several research studies in recent years, the translation of radiomics into clinical practice has been rather slow. For example, only a small proportion of FDA-approved medical devices based on artificial intelligence involve radiomics.² As for clinical trials, a search on clinicaltrail.gov with the keywords “oncology” and “radiomics” yields 238 results, of which 40 completed, but only one reported the results of the study.³

Besides the inherent slowness of technological translation that is characteristic of any scientific discipline, several specific reasons for the scarcity of clinical applications of radiomic research could be identified. These include poor standardization of imaging protocols and radiomic analysis, a lack of proper external validation of trained machine-learning models, and a missing focus on the interpretability and biological meaning of identified radiomics features.⁴

From a technical perspective, radiomics encompasses multiple steps, including imaging acquisition and preprocessing, segmentation, feature extraction and selection, modeling, and evaluation. This complexity has consequences on the generalizability of individual studies, because of the different approaches that could be followed at each step, resulting in heterogeneous and sometime conflicting results.^{5, 6} Collaborative efforts are underway to address the challenges and the sources of bias associated with radiomics and have been translated into updated reference standards, reporting guidelines and quality assessment tools.^6-8 As radiomics is a rapidly developing field, it is natural that such guidelines promoted by scientific societies and international collaborations remain current and aligned with advancements in radiomics technology.

Validation of radiomic models on an external dataset is a milestone for testing their clinical utility. The availability of more high-quality datasets from institutions for use as test sets for radiomic models facilitates greater confidence and validation of radiomic findings.⁹ Inspired by findability, accessibility, interoperability, and reusability principles and Open Science initiatives, most recent radiomic guidelines encourage research teams to share the codes and data produced during their studies, with the aim of improving the transparency and reproducibility of the results.

The integration of radiomics into the existing infrastructure of diagnosis and reporting is another essential step for maximizing its utility and impact in real-world medical settings. It is imperative to ensure accessibility to end-users, particularly radiologists and oncologists. This requires the development of specialized analysis software incorporated into clinical workflows that can interact with reporting systems.¹⁰

Imaging features capture macroscopic characteristics of the tumor, but the link with histological, genetic or molecular features is not always assured. This can hamper the interpretability of radiomic models. Clinical translation of radiomic studies could succeed by bridging the gap between such “radiomic signature” and the underlying pathology. To achieve this, establishing a tissue-based pathological validation of a radiomic model may lead to a deeper understanding of the relationship between medical imaging data and pathological features of the tumor.⁴

Radiomics provides radio-phenotypic characteristics which are useful for individual patient management. However, in order to become part of personalized medicine like other omics have done (e.g. genomics), prospective multicenter radiomic studies that incorporate the radiomic signatures into clinical trials as primary or secondary endpoints should be conducted, following established processes of standardization and validation.^{2, 4-6} To assess the clinical utility of a radiomic model, its performance in the intended clinical setting should be measured and the risk-benefit balance for the patient should be evaluated, especially because it may involve changes in clinical decisions. For a radiomic model to be clinically useful, it must lead to either superior clinical outcomes compared with standard options, noninferior outcomes with reduced risks, or avoidance of unnecessary or ineffective treatments. That said, clinical utility of a radiomic model is absent if it only identifies statistically different patient groups without impacting recommended clinical management.²

Another crucial aspect for advancing the implementation of radiomics in clinical practice is the education of medical students and radiologists in informatics, coupled with collaboration with computer scientists. Universities and research institutes play a pivotal role in driving this change by offering interdisciplinary programs and fostering collaborative research initiatives. Incorporating informatics education into medical and radiology curricula equips future healthcare professionals with the necessary skills to effectively navigate and analyze radiomic data.

While radiomics holds immense promise for revolutionizing personalized patient care, its effective integration into medical imaging and clinical practice necessitates concerted efforts to address technical, standardization, reproducibility, and interpretability challenges. By fostering collaboration, transparency, and validation, radiomics can pave the way for more precise diagnosis, treatment, and management of diseases, ultimately improving patient outcomes, and advancing healthcare delivery.

在过去的十年中，放射组学彻底改变了医学成像文献，它可以从磁共振图像和其他成像模式中提取大量定量数据，用于各种诊断和预后目的，特别是在肿瘤学方面。放射组学可以捕捉肿瘤的成像特征，这些特征超出了传统的视觉分析，并且与组织学类型、基因突变、治疗反应和患者生存密切相关。¹

尽管近年来多项研究取得了可喜的成果，但放射组学向临床实践的转化却相当缓慢。例如，FDA批准的基于人工智能的医疗设备中只有一小部分涉及放射组学。²至于临床试验，在 Clinicaltrail.gov 上用关键词“oncology”和“radiomics”进行搜索，得到 238 个结果，其中 40 个已完成，但只有一个报告了研究结果。³

除了任何科学学科所特有的技术转化速度缓慢之外，还可以确定放射组学研究临床应用稀缺的几个具体原因。其中包括成像协议和放射组学分析的标准化较差，缺乏对训练有素的机器学习模型的适当外部验证，以及缺乏对已识别放射组学特征的可解释性和生物学意义的关注。⁴

从技术角度来看，放射组学包含多个步骤，包括图像采集和预处理、分割、特征提取和选择、建模和评估。这种复杂性会对个别研究的普遍性产生影响，因为每个步骤可能采用不同的方法，从而导致异质性甚至有时相互矛盾的结果。^{5, 6}各方正在共同努力解决与放射组学相关的挑战和偏差来源，并已转化为更新的参考标准、报告指南和质量评估工具。^6-8由于放射组学是一个快速发展的领域，科学协会和国际合作推动的此类指南自然会保持最新状态并与放射组学技术的进步保持一致。

在外部数据集上验证放射组学模型是测试其临床实用性的一个里程碑。来自机构的更多高质量数据集可用作放射组学模型的测试集，有助于提高放射组学研究结果的可信度和验证。⁹受可查找性、可访问性、互操作性和可重用性原则以及开放科学倡议的启发，最新的放射组学指南鼓励研究团队共享研究过程中产生的代码和数据，目的是提高结果的透明度和可重复性。

将放射组学整合到现有的诊断和报告基础设施中是最大化其在现实世界医疗环境中的效用和影响的另一个重要步骤。必须确保最终用户，特别是放射科医生和肿瘤科医生的可访问性。这需要开发能够与报告系统交互的临床工作流程中的专用分析软件。¹⁰

成像特征捕捉肿瘤的宏观特征，但与组织学、遗传或分子特征的联系并不总是确定。这可能会妨碍放射组学模型的可解释性。通过弥合这种“放射学特征”与潜在病理学之间的差距，放射组学研究的临床转化可以取得成功。为了实现这一目标，建立基于组织的放射组学模型的病理验证可能会导致更深入地了解医学成像数据与肿瘤病理特征之间的关系。⁴

放射组学提供放射表型特征，这对于个体患者管理很有用。然而，为了像其他组学（例如基因组学）那样成为个性化医疗的一部分，应按照既定的标准化和验证流程进行前瞻性多中心放射组学研究，将放射组学特征纳入临床试验作为主要或次要终点。^{2, 4-6}为了评估放射组学模型的临床效用，应测量其在预期临床环境中的表现，并评估患者的风险收益平衡，特别是因为它可能涉及临床决策的变化。为了使放射组学模型在临床上有用，它必须导致与标准选择相比更好的临床结果、降低风险的非劣质结果，或者避免不必要或无效的治疗。也就是说，如果放射组学模型仅识别统计上不同的患者组而不影响推荐的临床管理，则它的临床效用是不存在的。²

推进放射组学在临床实践中实施的另一个重要方面是对医学生和放射科医生进行信息学教育，以及与计算机科学家的合作。大学和研究机构通过提供跨学科项目和促进合作研究举措，在推动这一变革方面发挥着关键作用。将信息学教育纳入医学和放射学课程，为未来的医疗保健专业人员提供有效导航和分析放射组学数据的必要技能。

虽然放射组学在彻底改变个性化患者护理方面有着巨大的前景，但其有效融入医学成像和临床实践需要共同努力解决技术、标准化、可重复性和可解释性挑战。通过促进协作、透明度和验证，放射组学可以为更精确的疾病诊断、治疗和管理铺平道路，最终改善患者的治疗结果并推进医疗保健服务。