ICNN-Ensemble: An Improved Convolutional Neural Network Ensemble Model for Medical Image Classification

Abstract

Deep learning (DL) classification has become a major research topic in the areas of cancer prediction, image cell classification, and image classification in medicine. Furthermore, DL classification is the core of other subfields. Owing to various forms of ensemble models, DL models have achieved state-of-the-art performances in fields such as medicine. However, the existing models cannot solve the problem of generalization perfectly and proposed solutions only for tasks with specific datasets. Most state-of-the-art classification models presented their results in ImageNet dataset, and models elaborate on the insights of the dataset. Nonetheless, model architectures or pretrained models cannot provide the same accurate results for datasets with different classes than ImageNet. Hence, this research proposes an improved convolutional neural network ensemble (ICNN-Ensemble) based on the representation of high-resolution image channels (RHRIC) and a systematic model dropout ensemble (SMDE). ICNN-Ensemble exploits image channels after applying RHRIC and RGB images in their original forms, which accesses more residual feature connections and represents more insight into image channels. Furthermore, SMDE is applied to choose ensemble members, considering the changes of the accurate prediction field (APF) in the ICNN-Ensemble model. In addition, the proposed model executes ensembling during the test set prediction, which allows the model to be trained with larger batches and images compared to ensemble model's final results during training, allowing maximal effective usage of the graphics processing unit (GPU). Despite the small size of the model, the results of benchmarking for Malaria cell images dataset clearly illustrated that the ICNN-Ensemble model achieved significantly more accurate results than other base state-of-the-art models.