Deep learning techniques for clinical decision support

临床医生越来越迫切地需要使用扫描仪产生的大量数据，从低级别x射线到高分辨率多维CT扫描。分析这些数据可能是一个主观的过程，对临床医生来说也很耗时。基于强大的人工智能方法的系统可以大大提高临床估计和决策的准确性。虽然该项目将专注于在人工智能的一般领域，特别是医学成像领域推进计算机科学方法，但临床应用仍将在视线范围内，临床投入将从纳菲尔德骨科医院的合作者那里寻求。我们将利用有注释和原始（未注释）数据的可用性，可能来自成人髋关节的MRI扫描，以设计可以支持临床决策的方法。这些方法将在扫描中预测解剖特征的存在或缺失，并将定位特定的关键点，从而可以在临床中自动进行某些测量（如角度）。语义分割技术还可以帮助识别特定感兴趣区域周围的轮廓，以及帮助临床医生分析这些区域内的扫描仪信号。一般目标与分割和地标识别有关。更具体地说，将特别注意预测的稳健性。这些代理人很少能量化他们对自己预测的信心。因此，我们的目标是计算并输出系统自身结果的不确定性，并允许临床医生最终决定是否接受或调整该辅助系统的任何输出。这些方法将支持安全筛选和监测出现长期疾病的患者，如早期骨关节炎。该系统最终可以帮助临床医生决定是否采用保守治疗（如物理治疗）或进行更具侵入性的干预（如髋关节置换术）。推迟髋关节置换术将大大减少所谓的“翻修”的需要，这是第二次髋关节置换术，当第一个假髋关节磨损时进行。修订极其昂贵和复杂，因此帮助决策和以安全的方式推迟一些干预可能会为NHS节省大量成本。该项目属于EPSRC医疗技术研究领域，特别是实施“适应不确定性的复杂模型和决策支持系统”。一旦得到验证，建议的方法将成为临床医生选择和证明所选患者特异性治疗的辅助手段。我们的临床合作者是牛津大学纳菲尔德骨科中心的Sion Glyn-Jones教授和Simon Newman教授。

Clinicians are increasingly pressed to use vast volumes of data generated by scanners, ranging from low grade xrays to high resolution multi-dimensional CT scans. Analysing this data can be a subjective process, and is also time consuming for a clinician. A system based around robust artificial intelligence methods can greatly improve the accuracy of estimates and decisions made in the clinic.Whilst this project will focus on advancing computer science methods within the general area of artificial intelligence in general, and medical imaging in particular, the clinical applicability will remain in sight, and clinical input will be sought from collaborators at the Nuffield Orthopaedic Hospital.We will leverage the availability of both annotated and raw (unannotated) data, likely from MRI scans of adult hips in order to design methods which can support clinical decisions. These methods will predict the presence or absence of anatomical features within the scan, and will localise specific key points which can automate certain measurements (such as angles) taken in the clinic. Semantic segmentation techniques can also help identify contours around specific areas of interest, as well as help clinicians analyse the scanner signal within such areas.The generic aims are related to the segmentation and landmark identification in general. More particularly, special attention will be given to the robustness of the prediction. Such agents can rarely quantify their confidence in their prediction. As such, we will aim to calculate and output a measure of the uncertainty of the system in its own results, and will allow the clinician have final say in whether to accept or adjust any output from this adjunct system.These methods will support safe screening and monitoring patients with emerging long-term conditions such as the early stages of osteoarthritis. The system could ultimately help the clinician decide whether to employ conservative management (e.g. physiotherapy) or to make a more invasive intervention (such as hip replacement).Delaying hip replacement will greatly reduce the need for what is known as a "revision", which is a second hip replacement operation, carried out when the first prosthetic hip has worn out. Revisions are extremely expensive and complicated, so aiding the decision making and delaying some intervention in safe ways will potentially save the NHS significant costs.This project falls within the EPSRC Healthcare Technologies research area, particularly implementing "Complex models and decision-support systems that accommodate uncertainty". Once validated, the proposed methods will be an adjunct to clinicians in choosing and justifying the chosen patient-specific treatment.Our clinical collaborators are Professors Sion Glyn-Jones and Simon Newman from the Nuffield Orthopaedic Centre in Oxford.