VGG16 CLASSIFICATION OF MICROSATELLITE INSTABILITY IN COLORECTAL CANCER BY DEEP LEARNING

Abstract

Recently, with the development of technology, deep learning method, which is a branch of artificial intelligence, has become increasingly widespread in the field of health, with its use in the diagnosis, classification and determination of appropriate treatment. This study, which was carried out with this method, is important in determining the most appropriate treatment option for the patient, such as early estimation of microsatellite instability (MSI) in colorectal cancer patients, determination of appropriate treatment, reduction of undesirable side effects, and prevention of time and cost for the disease. Method: In this study, the VGG16 model was developed using deep learning techniques, which are convolutional neural network and transfer learning methods, and the classification of microsatellite instability was provided. Using the colorectal cancer hematoxylin and eosin (H&E) stained 150000 histopathological image dataset, which is open access via the Kaggle website, 80% was reserved for training and 20% was reserved for testing. With these separated data sets, the training of the deep learning-based VGG16 model was carried out using high graphics processing units (GPU) over Google Colab, a free cloud environment. Results: For the training of the model, the epoch, learning rate, momentum coefficient and group size hyperparameters were determined as 10, 0.00001, 0.9 and 64, respectively, to classify microsatellite instability in colorectal cancer. With the proposed model VGG16, classification performance accuracy 89.4%, precision 92.9%, sensitivity 85.3% and AUC 89.4% were obtained. Conclusion: In line with these results, it is believed that it will help pathologists to make computer-aided decisions in the clinical setting.