基於特徵結構性學習之螢光顯微鏡圖片癌細胞辨識系統

Abstract

在這篇論文裡，我們提出一套自動化的癌細胞辨識系統，用於在螢光顯微鏡圖片上偵測並辨識出目標癌症細胞。完整的流程包含了四個主要部分：圖片前處理、候選者偵測、候選者篩選以及分類。不同於現有的做法只針對細胞圖片進行各種特徵設計並加強分類器的訓練，我們加入特徵屬性的概念來完成一個結構性學習的模型訓練。在我們的做法中，我們將視覺上重要的資訊轉換成不同特徵屬性的設計幫助模型進行分類，由此我們可以進一步學習出一般特徵與特徵屬性之間相輔相成的關係，並且使我們的分類模型能夠處理變異度極高的細胞形態。除此之外，在我們的實驗中，有限或不均衡的資料分布造成模型在分類上的困難度大幅增加，這個現象也經常出現在其他生醫影像的應用中，因此針對這個問題我們使用級聯支持向量機器架構並給予不同數量群組不同的權重。在對於一張測試影像進行辨識過程時，我們首先會在影像上找出可能候選者的集合，再將這些候選者逕行篩選與分類，而最後系統的輸出是一個依照癌症細胞可能性的排序。實驗結果顯示，我們的結構性模型能夠提供有效並實用的結果。

Automatic cell identification is an important step in bioimaging informatics. In this thesis, we propose an automatic cancer cell identification system for fluorescence microscopy images based on a structural learning cell model. The system consists of four main steps: image pre-processing, putative candidate region detection, candidate post-processing, and structural inference by a trained cell model. Unlike existing approaches aiming at feature or descriptor design, we introduce ideas of semantically rich attribute and propose an algorithm to learn structural relationship between ordinary hand-crafted features and attributes. The structural model has better ability to deal with large variations of cell morphology. However, another issue in our implementation is limited and imbalanced data, which is a common problem in biological image research. To tackle the problem, we train the structural cell model within a cascaded support vector machine framework using class-weighting scheme. During testing process, a set of candidates will be detected on the image and send to the structural model. The output of overall system is a ranking list ordering by possibility of each candidate being a cancer cell. Simulation results demonstrate the efficiency and effectiveness of our proposed method.