Automatic Diagnosis Support System Using Nuclear and Luminal Features
Yuriko Harai, Toshiyuki Tanaka · 2015
We present a method of automatic colorectal cancer diagnosis that can quantify cellular and structural tissue information. In this paper, we consider sixteen-dimensional features, consisting of the nuclei-cytoplasm (NC) ratio, connected nuclei area, and atypical lumen ratio. For the purpose of imitating the conditions of accurate medical diagnosing, we introduce a four-class classification for group 1, group 3 low, group 3 high, and group 5 biopsies (group 5 biopsies include well-, moderately, and poorly differentiated) in contrast to most previous works proposed in the literature, which classify biopsies into two or three classes. The image set used in this paper consists of 400 images stained from 123 patients by hematoxylin and eosin (the HE method). We compared the performance of the proposed method with a method using texture features that have been widely used in previous studies. Two classification tests were performed, leave-one-ROI-out cross-validation (CV) and leave-one-specimen-out CV. As a result, the proposed method obtained a classification accuracy of 95.0% for ROI-based CV and 78.3% for specimen-based CV.