Histopathological Image and Lymphoma Image Classification using customized Deep Learning models and different optimization algorithms

Medical image classification is a challenging task that is time-consuming, requires deep domain expertise, and is also prone to error. Automated image classification can be used to complement the domain experts for addressing the abovementioned concerns in medical image classification. In this paper, we have used the KIMIA Path 960 dataset, and the NIA curated dataset for image classification. KIMIA Path 960 dataset has 960 images belonging to 20 classes. NIA curated dataset contains lymphoma images belonging to 3 classes. We have used pre-trained Resnet50 model and customized it with adding layers to the pre-trained models for this dataset. In another method, we have used a five-layer CNN for classification. We have also done experiments with different optimization algorithms, namely Adam, AdaMax, AdamW, and Radam, with all the above-mentioned models. Our method using the KIMIAPath 960 dataset and the AdamW optimization algorithm has achieved an accuracy of 99.9%, which is better than the accuracies obtained using Inception Recurrent Residual Convolutional Neural Network (IRRCNN) (98.79%) and Bag of Visual Words (96.5%) approach in other papers. We have also used the same method in the NIA curated dataset of lymphoma images and have achieved the best accuracy of 98.13%, which is better than the papers using the same dataset. We have also observed that the accuracy of the deep neural network is dependent both on the neural network architecture as well as the optimization algorithm used in the neural network architecture.