利用圖形處理器計算大規模基因關聯分析暨病理切片分析

Abstract

本論文是由兩個研究主題構成，兩個主題皆以圖形處理器分析大量醫學資料。第一部分：本研究以計算基因組關聯分析為中心，發展出一個新的演算法，並實作在圖形處理器上。基因組關聯分析廣泛的被運用在研究基因組和性狀間的關係。本研究在圖形處理器上之實作可以大幅加速基因組關聯分析的運算，並且擁有極佳的可擴展性。本研究的核心在加速基因組關聯分析中最耗費時間的 P 值計算。利用此新演算法加速的圖形處理器實作程式將提供基因研究員一個更好的工具，在數分鐘內能完成數以百億組基因組關聯分析，並且由於其擁有高度可擴展性，此演算法能夠在多片圖形處理器系統上計算更龐大的資料。第二部分：病理切片是從病人患病處切下之組織細胞所做成的切片，病理切片在癌症診斷中佔據著非常重要的地位，然而由於病理切片沒有特定形狀、顏色，加上其判斷規則之繁複以及影像之大、資料收集之困難，如何讓電腦自動判斷病理切片至今仍是一項困難的挑戰。本研究利用低解析度的影像片段作為判斷標的，利用深度學習最後達成在判斷前列腺病理切片格里森分數 3+3 及 4+4 有 95% 的準確率。

This thesis is a combination of two pieces of research; both make use of the Graphics Processing Unit (GPU) for calculation and both deals with large medical data. Part I. We develop a fast algorithm as long as a CUDA code for GWAS (Genome-Wide Associate Studies) to find the relation between genomes and targeted traits. This algorithm can work efficiently on GPU and has high scalability. The core of the algorithm is an accurate and fast p-value calculating method, which accelerates the most time-consuming part of GWAS problems. With the algorithm, researchers can now deal with tens of billions of SNP to trait pair in only a few minutes. Even better, since this algorithm is highly scalable, users can increase the problem size as long as one has enough computing power. Part II. Pathology images are whole slide images of patient tissues, and such images provide one critical tool for cancer diagnosis. Despite its importance, how artificial intelligence can automatic the Gleason scores grading remain a challenge. The challenge is due to the large size of digital whole slide image (around giga-pixels), the variation of stained color, cell texture,..., and limited labeled data. We propose a low-resolution method instead of existed patch-based methods and achieved an accuracy of 95 percent on classifying prostate pathology images with 3+3 or 4+4 Gleason scores.