Next-generation imaging biomarkers in neuro-oncology using artificial intelligence: overcoming key challenges towards clinically applicable AI

在过去的几十年里，在改善癌症患者的治疗和结果方面取得了巨大进展。然而，脑肿瘤是该提案的重点，由于其巨大的异质性和潜在的复杂生物学，其预后仍然主要较差。然而，希望在于个体化和分子靶向治疗方法。在这种情况下，开发准确和广泛适用的生物标志物来评估这些新疗法的疗效变得至关重要。磁共振成像（MRI）在这方面特别重要，人工智能（AI）领域的最新进展表明，放射图像数据的定量分析取得了显着进展。在这个项目中，我们将在SPP-2177第一个资助期的基础上进一步扩大我们的关键发展，目标是缩小在脑肿瘤成像领域实施临床适用AI的差距。通过利用神经肿瘤学中的大规模多模态数据资源（具有纵向多参数MRI、分子和临床数据），其中超过5000例患者来自先前进行的神经肿瘤前瞻性多中心临床研究，我们的项目将解决以下关键目标：（1）为进一步提高性能，我们以前开发的国家的普遍性和临床实用性，用于脑肿瘤分割模型的人工智能模型，用于通过实施时间一致、不确定性感知和持续学习模型进行定量肿瘤反应评估;（2）实施新的隐私保护人工智能技术，这将允许在使用多机构数据开发基于不同人群样本的人工智能模型时克服机构之间的数据共享需求;以及（3）实现AI模型的可解释预测，从而解决基于AI的分类模型的“黑箱”性质。总之，预期的发展不仅将解决在脑肿瘤成像领域实现临床适用AI的关键挑战，还将成为AI在放射学领域有意义应用的蓝图。

Over the past decades, tremendous progress has been made in improving cancer patient treatment and outcomes. Brain tumors, which are the focus of this proposal, are however still associated with a predominantly poor prognosis due to their enormous heterogeneity and underlying complex biology. However, hope lies in individualized and molecularly targeted treatment approaches. In this context, it has become critical to develop accurate and broadly applicable biomarkers to assess the efficacy of these novel therapies. Magnetic resonance imaging (MRI) is of particular importance in this regard, and recent advances in the field of artificial intelligence (AI) have demonstrated remarkable progress in the quantitative analysis of radiological image data. In this project, we will build on and further expand our key developments from the first funding period of SPP-2177, with the goal of closing the gap towards implementing clinically applicable AI in the field of brain tumor imaging. By leveraging a large-scale multimodal data resource in neuro-oncology (with longitudinal multiparametric MRI, molecular and clinical data) with more than 5000 patients from previously conducted prospective multicenter clinical studies in neuro-oncology, the following key-objectives will be address within our project: (1) to further improve the performance, generalizability and clinical utility of our previously developed state-of-the-art AI models for brain tumor segmentation model for quantitative tumor response assessment through implementation of temporally consistent, uncertainty aware and continual learning models; (2) to implement novel privacy-preserving AI techniques which will allow to overcome the need of data-sharing between institutions when using multi-institutional data for the development of AI models based on diverse population samples; and (3) to implement interpretable predictions of AI models and thereby addressing the “black-box” nature of AI based classification models. In summary, the anticipated developments will not only address the key challenges towards achieving clinically applicable AI in the field of brain tumor imaging, but also serve as a blueprint for the meaningful application of AI in the field of radiology.