Machine learning for systems pathology

Medical informatics--Statistical methods;Cancer invasiveness--Measurement--Statistical methods;Machine learning;Pathology--Statistical methods;Ovaries--Tumors--Diagnosis--Statistical methods;System theory

Verleyen, Wim ; Stokes, Anne (V. Anne) ; Stokes, Anne (V. Anne)

University:University of St Andrews

Department:Biology (School of)

关键词: Medical informatics--Statistical methods;    Cancer invasiveness--Measurement--Statistical methods;    Machine learning;    Pathology--Statistical methods;    Ovaries--Tumors--Diagnosis--Statistical methods;    System theory;   

Others  :  https://research-repository.st-andrews.ac.uk/bitstream/handle/10023/4512/WimVerleyenPhDThesis.pdf?sequence=3&isAllowed=y

来源: DR-NTU

【 摘 要 】

Systems pathology attempts to introduce more holistic approachestowards pathology and attempts to integrate clinicopathologicalinformation with “-omics” technology. This doctorate researchestwo examples of a systems approach for pathology: (1) apersonalized patient output prediction for ovarian cancer and (2) an analyticalapproach differentiates between individual and collective tumourinvasion.During the personalized patient output prediction for ovariancancer study, clinicopathological measurements and proteomicbiomarkers are analysed with a set of newly engineered bioinformatictools. These tools are based upon feature selection, survivalanalysis with Cox proportional hazards regression, and a novelMonte Carlo approach. Clinical and pathological data proves tohave highly significant information content, as expected; however,molecular data has little information content alone, and is only significantwhen selected most-informative variables are placed in thecontext of the patient’s clinical and pathological measures. Furthermore,classifiers based on support vector machines (SVMs) thatpredict one-year PFS and three-year OS with high accuracy, showhow the addition of carefully selected molecular measures to clinicaland pathological knowledge can enable personalized prognosispredictions. Finally, the high-performance of these classifiers arevalidated on an additional data set.A second study, an analytical approach differentiates betweenindividual and collective tumour invasion, analyses a set of morphologicalmeasures. These morphological measurements are collectedwith a newly developed process using automated imaging analysisfor data collection in combination with a Bayesian network analysisto probabilistically connect morphological variables with tumour invasionmodes. Between an individual and collective invasion mode,cell-cell contact is the most discriminating morphological feature.Smaller invading groups were typified by smoother cellular surfacesthan those invading collectively in larger groups. Interestingly,elongation was evident in all invading cell groups and was not aspecific feature of single cell invasion as a surrogate of epithelialmesenchymaltransition. In conclusion, the combination of automatedimaging analysis and Bayesian network analysis providesan insight into morphological variables associated with transitionof cancer cells between invasion modes. We show that only twomorphologically distinct modes of invasion exist.The two studies performed in this thesis illustrate the potential ofa systems approach for pathology and illustrate the need of quantitativeapproaches in order to reveal the system behind pathology.