Deep Learning Approaches for Brain Tumor Diagnosis using Fused Layer Accelerator

Abstract

Deep Convolutional Neural Networks (DCNNs) are an emerging field in biomedical processing. Tumor classification is a key stage in the pathology analysis process, and deep learning algorithms for brain tumor categorization have recently shown promising results. However, these approaches often require more storage and more expensive training procedures to input a large number of training variables. To address this issue, light-weight deep learning models should be investigated without reducing classification accuracy. The aim of this study was to compare the classification rate of three pre-trained Transfer Learning classifiers, namely InceptionResNetV2, EfficientNetV1, and MobileNetV2, in categorizing brain tumors into four classes such as glioma tumors, meningioma tumors, pituitary tumors, and normal patients. In this article, attention modules based on pretrained deep learning models such as MobileNetV2, EfficientNetV1, and InceptionResNetV2 were highlighted. Following the fully connected layers and the ReLU6 layer, attention and convolution modules were integrated to obtain high-level object-based and critical semantic information. The effectiveness of this strategy was demonstrated by building a CNN accelerator based on the fusion of the top five convolutional layers of MobileNetV2, EfficientNetV1, and InceptionResNetV2 networks and comparing it to a Python accelerator. The EfficientNetV1 model showed the best results compared to the InceptionResNetV2 model and MobileNetV2 model.