Abstract
In this research, we conducted to generate Convolutional Neural Network (CNN) models which can distinguish between cancer cells and normal cells in the large intestine under 11 input image size (number of pixels) conditions (50×50 px, 100×100 px, 200×200 px, 300×300 px, 400×400 px, 500×500 px, 600×600 px, 700×700 px, 800×800 px, 900×900 px, 1000×1000 px) for the purpose of clarifying the effect of the input image size on the cancer cells identification in colorectal cancer pathology image. When we conducted to train models, we calculated model accuracy by 5-fold cross validation in each condition. As a result, the accuracy rate increased as input image size increased in 3 conditions (50×50 px, 100×100 px, 200×200 px). However, the accuracy rate of the model trained by 4 conditions (700×700 px, 800×800 px, 900×900 px, 1000×1000 px) were lower than that trained by other conditions. These results are thought to be caused by setting the kernel size to the same value in all conditions. Therefore, we concluded that it is necessary to determine an appropriate kernel size and stride for each input image size which can extracted features included in the input image when we conduct to generate a model for the purpose of distinguish between cancer cells and normal cells in the large intestine.
