Machine learning for improving blood cancer diagnostics via flow cytometry

设在皇家利物浦医院的综合血液肿瘤学诊断服务（HODS）为默西塞德郡和柴郡癌症网络进行血癌检测，包括淋巴瘤、骨髓瘤和急性白血病。其中一种检测方法是使用流式细胞术检测血液、骨髓或其他体液（如脑脊液）制备的细胞悬液中的细胞。流式细胞术以多通道模式（例如8或12通道）进行，其中每个通道报告不同的特性：群体中每个细胞的大小、形状、荧光表达强度（通过抗体结合等替代细胞表面上蛋白质表达的存在/水平）。目前在HODS的做法是人工检查数据，例如通过对细胞不同性质的双参数点图进行可视化，以确定细胞群体的免疫表型并枚举任何异常/恶性亚克隆。样本内的群体根据其表面抗原表达进行分组，具有与世卫组织分类的不同淋巴瘤或白血病相对应的某些表达模式。然后，专家分析师查看现有软件生成的图并撰写报告，临床医生使用该报告做出最终诊断。目前的实践是劳动密集型的，并且有可能错过细胞表型参数之间的非线性或复杂关联。此外，使用传统方法很难确定低参与程度。这些数据非常适合机器学习的应用（例如随机森林或人工神经网络），因为在本地数据库中存储和标记了丰富的数据（可追溯到10年前），并且这些方法在识别人眼不明显可见的数据中的微妙模式方面具有明显的潜力。一种新系统很快将在实验室上线，其中罕见事件检测将是关键，例如，在病人接受治疗（如化疗）后识别少量癌细胞（例如，每百万几个细胞）。优化分析所需的信息学方法有很大的创新空间。学生将接受机器学习、信号检测和软件工程方法方面的全面培训，并将能够定期访问免疫表型实验室，了解协议和程序，并了解提高诊断能力的最关键因素。该项目将开发用于癌症类型识别和罕见事件检测的机器学习方法。他们接下来将开发新的软件，以可视化数据和运行机器学习模型，以补充当前实践的方式用于HODS。

The integrated Haemato-oncology diagnostics service (HODS) based at the Royal Liverpool Hospital performs testing for the Merseyside and Cheshire Cancer Network for blood cancers including Lymphoma, Myeloma and Acute Leukaemia. One of tests performed is to use flow cytometry to examine the cells within a cell suspension prepared from blood, bone marrow or other body fluid e.g. CSF. Flow cytometry is performed in multi-channel mode (e.g. 8 or 12 channels) where each channel reports a different property: size, shape, intensity of fluorochrome expression (a surrogate for presence/level of protein expression on the cell surface via antibody binding etc) of each cell in the population. Present practice at HODS is for manual examination of data, for example via visualisation of dual parameter dot-plots of different properties for cells, to both determine the immunophenotype of the cellular populations and enumerate any abnormal/malignant sub clones. Populations within the sample are grouped according to their surface antigen expression, with certain patterns of expression corresponding to the different WHO classified lymphomas or leukaemias. An expert analyst then views the plots produced by existing software and writes a report, which is used by the clinician to make a final diagnosis. The current practice is labour intensive and has the potential to miss non-linear or complex associations between phenotypic parameters of cells. In addition identification of low levels of involvement is difficult using conventional methods.The data is ideally suited to the application of machine learning (e.g. random forest or artificial neural networks), due to the wealth of data stored and labelled in a local database (going back >10 years), and clear potential for such methods to identify subtle patterns in the data not obviously visible to the human eye. A new system is soon to come online in the laboratory, in which rare event detection will be key, for example to identify small counts of cancerous cells (e.g. a few cells per million) following a patient undergoing treatment, such as chemotherapy. There is significant room for innovation in the informatics approach needed to optimise analysis.The student will be fully trained in machine learning, signal detection and software engineering methods, and will be able to get regular access to the immunophenotyping lab to learn about protocols and procedures, and understand the most critical factors in developing improved diagnostic ability. The project will develop ML methods for identification of cancer types, and rare event detection. They will next develop new software, to visualise data and run the ML models, to be used in HODS, in a complementary manner to current practice.