C4: Biomedical Imaging I

REGION SEGMENTATION AND FEATURE EXTRACTION IN GASTRIC BIOPSY IMAGES FOR AUTOMATIC DIAGNOSIS SUPPORT SYSTEM

Emi Morotomi, Toshiyuki Tanaka

Keio University, Japan

Recently, the shortage of pathologists becomes serious problem in Japan, while the number of cancer patients has been increasing year by year. Therefore, it is required to construct pathological diagnosis support system. Specially, it is important that all the cases are classified into two groups: neoplastic lesion and nonneoplastic lesion. In this paper, we propose a method with image analysis to classify the gastric tumors into neoplastic lesion and non-neoplastic lesion.

Our proposed method mainly consists of image input, region extraction, feature calculation, and discriminant analysis. This study focuses on the region extraction in the whole flow. Our method has gland extraction method with a modified watershed algorithm, because pathologists take notice of glandular shapes and structures. We obtain several features from extracted regions, and perform discriminant analysis. By performing these processing, we try to improve the classification accuracy.

In order to extract glandular region, we use markercontrolled watershed segmentation. This method requires to set target domain marker and background domain marker. we set target domain marker based on extraction of lumen, and set background domain marker based on rough shape extraction of gland. The proposed method of lumen extraction contains removal of stroma, which is likely to be mistaken as lumen in the previous study.

Next, we compute features of obtained glandular shapes and structures. Several effective features of shape and structure is obtained by calculating the degree of separation. The results of our method shows that the classification ratio is approximately 4.5% better than the previous method in an experiment using 45 gastric biopsy images.  In the future, we will improve the accuracy of our algorithm by devising marker processing and introducing new effective features.

Organised by

Endorsed by

 

        Supported by
 

 

                          
     

       

                             
           
 

        Supporting Media