Improved Medical Image Segmentation Algorithm Using Two-Step CNN Architecture

Abstract

Segmentation of glandular cells (glands) is the basic step in an automated cancer grading system. Existing methods for gland segmentation based on convolution neural network (CNN) apply combined datasets of both benign and malignant glands for deriving a single learning model. However, benign and malignant glands have very different structural appearances; so we thought the process will demonstrate better performance if we apply different CNN models for small-scale image sets. Therefore, in this paper, we propose a two-step CNN learning model that segments the glands by applying different data models according to the data type. Our proposed segmentation system comprises the following two steps: a CNN model for classification and a CNN model for segmentation. The first step is to classify whether the input image is benign or malignant glands using the slightly modified pre-trained VGG16 model. In the second step, the gland is segmented by applying one of two modified U-Net networks according to the classified results. Experimental results demonstrated that our two-step model has better segmentation performance compared to a single learning model for both benign and malignant data.