Machine learning for improved clinical decision making ahead of epilepsy surgery

博士项目目标：利用正常/异常弱监督信号分解在FDG-PET上自动检测和分割代谢病变，区分癫痫发作和癫痫扩散区域利用生物学和临床知识优化相关病变的检测项目描述：背景：癫痫是最常见的严重神经系统疾病，在英国有60万人受到影响（https://www.epilepsysociety.org.uk/about-epilepsy）。30%以上的癫痫患者出现耐药性癫痫发作。局灶性癫痫发作始于大脑的某一部分，如果可以确定癫痫区，手术可能是一种选择。术前评估可能需要10 - 10年，包括一系列诊断程序。成像，尤其是核磁共振成像，在这个过程中起着核心作用。[18F]氟脱氧葡萄糖位置发射断层扫描（FDG-PET）比MRI灵敏得多，在我们中心尤其发达。目前的临床影像学评估依赖于耗时和主观的视觉分析，甚至对专家来说也是具有挑战性的。局灶性皮质发育不良（FCDs）是大脑的小畸形，是难治性癫痫最常见的基础之一。通常仅通过核磁共振就能检测到，但15-30%的手术前患者为“核磁共振阴性”，在国家转诊中心，这一比例可能更高。在这些方面，FDG-PET在过去的五年中已经变得非常依赖：在植入颅内电极之前，有非常好的定位假设是必不可少的，因为这些是侵入性的，风险很小，最多只能检测到大脑的10%，如果电极放置在几毫米远的地方，癫痫发作区域可能无法检测到。深度学习方法已经在异常和病变的检测和表征方面获得了关注[1-3]。由于训练后速度很快，因此比传统的统计/机器学习方法[4-10]更适合在临床工作站实施。深度学习工作只专注于MRI分析，很大程度上忽略了FDG-PET的高产量，特别是在MRI阴性患者中。一项ML研究研究了FDG-PET，仅基于不对称而忽略MRI [11]；另一种结合了FDG-PET和MRI，但使用手工特征和支持向量机(svm) /基于补丁的分类，而不是深度学习[12]。本研究将证明FDG PET和MR定量联合分析在临床中的可行性和实用性。它将支持临床医生进行患者管理，并可能使更多的患者接受手术，这可能代表着显著的成本节约和对患者生活质量的积极影响。目标：1)开发癫痫患者FDG-PET扫描定量分析工具。2)与基于核磁共振的工具相结合3)对开发的工具进行前瞻性验证并获得临床医生反馈4)将非成像信息整合到图像分析中，如抑郁评分、脑电图数据和符号学（直接通过语义信息或视频脑电图）

Aim of the PhD Project:Automate metabolic lesion detection and segmentation on FDG-PET using normal/abnormal weakly-supervised signal decomposition Differentiating seizure-onset and seizure-spread areas Harness biological and clinical knowledge for optimised detection of relevant lesions Project description:Background: Epilepsy is the most common serious neurological condition, with >600,000 people affected in the UK (https://www.epilepsysociety.org.uk/about-epilepsy). Over 30% of people with epilepsy have medication-resistant seizures. Focal seizures start in one part of the brain, and surgery may be an option if the epileptogenic zone can be identified. Pre-surgical assessment can take >1yr and involves a range of diagnostic procedures. Imaging, particularly MRI, has a central role in this process. [18F]fluorodeoxyglucose position emission tomography (FDG-PET) is much more sensitive than MRI and particularly well developed in our Centre. Current clinical evaluation of imaging relies on time-consuming and subjective visual analysis and is challenging even for experts. Focal cortical dysplasias (FCDs) are small malformations of the brain and one of the most common substrates underlying refractory epilepsy. They can often be detected by MRI alone, but 15-30% of presurgical patients are "MRI-negative", with the proportion likely higher in national referral centres. In those, FDG-PET has become much relied upon over the past five years: it is essential to have very good localization hypotheses prior to implantation of intracranial electrodes, as these are invasive, carry a small risk, and can at best sample ~10% of brain, with seizure onset zones potentially undetected if electrodes are placed just a few mm away. Deep learning methods have gained traction for detecting and characterising abnormalities and lesions [1-3]. As they are fast once trained, they are more suitable than traditional statistical/machine learning methods [4-10] for implementation on clinical workstations. Deep learning work has exclusively focused on MRI analysis, largely ignoring the high yield of FDG-PET, especially in MRI-negative patients. One ML study has investigated FDG-PET, based on asymmetries alone and ignoring MRI [11]; another combined FDG-PET and MRI but used handcrafted features and Support Vector Machines (SVMs) / patch-based classification rather than deep learning [12]. This research will demonstrate the feasibility and usefulness of quantitative joint analysis of FDG PET and MR in the clinical setting. It will support clinicians with patient management and may enable more patients to have surgery, which could represent significant cost savings and a positive impact on patient quality of life. Goals: 1) Produce tools for quantitative analysis of FDG-PET scans of patients with epilepsy. 2) Combine with MR-based tools 3) Obtain prospective validation of the tools developed and obtain clinician feedback 4) Integrate non-imaging information into the image analysis, e.g. depression scores, EEG data, and semiology (via semantic information or ictal video-EEG directly)