DEEP TRANSFER LEARNING BASE MODELSWITHOPTUNA HYPERPARAMETER OPTIMIZATIONON COLORECTAL CANCER IMAGES CLASSIFICATION

Colorectal Cancer (CRC) is one of the forms of cancer and the second deadliest disease. Previously, some molecular tests, including next-generation sequencing has been used for this cancer classification. However, the required data do not always available to all cancer patients because of high cost and technical barriers. In recent development, Hematoxylin and eosin-stained (HE) biopsy slides are regularly available for colorectal cancer patients. Fortunately, rapid development has shown that objective biomarkers can be extracted from these images using Deep Learning (DL) approaches especially convolutional neural networks (CNN). This paper reports some experimental comparisons of transfer learning based on Deep Learning with CNN architecture as feature extractor to decompose nine classes of colorectal carcinoma slides images. The base of models of Transfer Learning includes MobileNet_V2, EfficientNet_b2, DenseNet201, Inception_V3, and Resnet18. All the base models were trained using Optuna hyperparameter optimization framework which has flexible, modular and easy to combined with transfer learning. This research has two main contributions, first, this research can contribute insight to medical expert and computer scientist related to the current state of development deep transfer learning approaches for histopathological images classification especially in colorectal carcinoma (CRC). The second contribution of this research is to recommend some CNN base models with hyperparameter optimization experimental framework development for decomposition of CRC feature extraction, parameter optimization and classification of slides images. The results show that highest accuracy is achieved by Resnet18 base models which has 99.99% on test data 1 and 98.16% on test data 2.