Methods for reliable machine learning with applications in medical imaging

这项研究计划涉及开发机器学习(ML)和模式识别方面的先进计算技术，以及应用于大规模真实世界医学成像检查的新的验证和评估技术，以追求医疗诊断中的下一代应用。将开发拟议的技术和ML评估方法，以创建有助于提高神经发育障碍儿童患者护理标准的诊断和疾病表征技术。开发的技术将通过对患有自闭症、注意力缺陷多动障碍(ADHD)等各种疾病的患者的MRI检查的大型数据集进行评估。这将包括在公开可用的MRI检查中对拟议的新型ML技术进行评估，以及将这些技术转化为基于常规临床成像检查的应用程序(波士顿儿童医院、哈佛医学院，我在那里担任研究助理)。这将支持对作为该提案的一部分开发的新提出的技术是否在现实的临床人群中发挥潜在作用的更大规模的评估。广泛的常规临床影像检查支持各种医疗疾病(多发性硬化症、脑瘫、神经纤维瘤病、精神分裂症等)对创造的技术的验证。*拥有多学科技术背景(计算、医学物理)和在跨学科医学研究(神经科学、生物医学工程)和许多新的研究途径方面的丰富经验，我处于独特的地位，能够在这项拟议的研究中取得成功。这项研究计划将涉及使用现有的模式识别技术从大数据集中提取感兴趣的测量，以及开发新的通用ML算法和验证方法，并在大量脑MRI检查中进行广泛评估。ML将被用来组合由模式识别技术提取的测量，以改进诊断和无序表征。这项建议将涉及使用现有技术，以及开发新的ML方法，这些方法建立在我先前工作的基础上，包括允许灵活的决策边界的公式，该边界随测试偏差设置而变化，为诊断测试提供优化。*将使用严格的统计验证技术，结果将尽可能用独立的数据集进行确认(自闭症、ADHD等)。将开发新的验证技术来评估样本大小和误差一致性问题。这些验证技术将是通用的，因此所有ML应用程序开发人员都可以使用它们。*****

This research program involves the development of advanced computational technologies in machine learning (ML) and pattern recognition along with novel validation and evaluation techniques applied to large-scale real world medical imaging examinations in pursuit of the next generation of applications in medical diagnostics. The proposed technological and ML evaluation methods will be developed towards the creation of diagnostic and disease characterization technologies that can help improve the standard of patient care for children with neurodevelopmental disorders. Technologies developed will be evaluated with large datasets of MRI examinations of patients with a variety of conditions such as autism, attention deficit hyperactivity disorder (ADHD) and more. This will include the evaluation of proposed novel ML technology on publicly available MRI examinations and the translation of those technologies to applications based on routine clinical imaging exams (Boston Children's Hospital, Harvard Medical School, where I hold an appointment as a Research Associate). This will support larger scale evaluation of whether novel proposed technologies developed as part of this proposal have a potential role to play in a realistic clinical population. The availability of extensive routine clinical imaging examinations supports the validation of created technologies across an assortment of medical disorders (multiple sclerosis, cerebral palsy, neurofibromatosis, schizophrenia and more).******With a multidisciplinary technical background (computation, medical physics) and extensive experience in interdisciplinary medical research (neuroscience, biomedical engineering) and many novel research avenues identified, I am uniquely positioned to succeed in this proposed research. This research program will involve the extraction of measurements of interest from large datasets using existing pattern recognition techniques along with the development of novel general purpose ML algorithms and validation approaches extensively assessed on large collections of brain MRI examinations. ML will be employed to combine the measurements extracted by the pattern recognition techniques to improve diagnostics and disorder characterization. This proposal will involve the use of existing techniques, as well as the development of novel ML methods that build on the pre-existing contributions of my work, including formulations which allow for a flexible decision boundary that varies with the test bias setting, providing optimizations for diagnostic testing.******Rigorous statistical validation techniques will be employed and results will be confirmed with independent datasets wherever possible (autism, ADHD, etc.). Novel validation techniques will be developed to assess sample size and error consistency issues. These validation techniques will be general-purpose and thus can be used by all ML application developers. *****