C4: Biomedical Imaging I


Motohiro Nagao, Toshiyuki Tanaka

Keio University, Japan

Although the number of cancer patients is on the increase, the number of pathologists is less than required. In addition to the shortage of pathological doctors, there is a gap of histopathological diagnosis skills among hospitals because it is often the case that only a single pathologist has to examine the whole set of biopsy samples in his/her hospital, or that a doctor who does not specialize in histopathology may study tissues. These issues can lead not only burdens of pathological doctors but also an inequality in medical service by hospitals. In order to solve these problems, there is a significant necessity of developing a computer-aided diagnosis system of cancer biopsy. In the previous work, an entire HE stained image is classified into a single class. However, in case it has two or more types of atypia in it, for example a sample image labeled as cancer with both cancerous and benign areas, features extracted from it will weaken, which causes to underestimate the malignancy. In this study, we divide a sample image into smaller rectangular patches and classify them into the histopathological classes by atypia so that we do not overlook local malignancy. We firstly perform an unsupervised recognition of unsuitable patches without a duct, the structure that indicates a degree of malignancy, based on Otsu’s binarization and a 3-by-3 Laplacian filter with a 10-by-10 averaging filter. Next, we input the patches to the former layers of Alex Net as a feature extractor. Then each set of extracted features is applied to a fully-connected neural network, outputting a likelihood of its class. The true positive rates (sensitivities) of three classes; Group1 (normal tissue), Group3 (adenoma) and Group5 (carcinoma) are more than 80% respectively.

Organised by

Endorsed by


        Supported by





        Supporting Media