Deployment of a Mobile Application Using a Novel Deep Neural Network and Advanced Pre-Trained Models for the Identification of Brain Tumours

Medical image processing is significant in clinical diagnosis and treatment. However, the traditional approach for examining these images has hit its performance limit, and requires a considerable amount of time and effort. In addition, the misdiagnosis of brain tumour types from these images can hinder patients from receiving proper medical treatment and can diminish their chances of survival. The emerging deep learning technique has shown good results in classification problems. In the current study, existing transfer learning models and a novel model were applied to classify augmented magnetic resonance images (n = 12,256 images). These methods were used to identify three different kinds of brain tumours from the images: meningioma, glioma, and pituitary tumours. Accuracies of 83.30%, 79.54%, 81.83%, 82.49%, 85.21%, and 91.73% were obtained using ADAM optimizers for VGG-16, VGG-19, Inception V3, Xception, Mobile net, and the proposed Lightweight Sequential net, respectively. Furthermore, we achieved accuracies of 50.45%, 60.24%, 57.85%, 61.98%, 75.14%, and 84.82% while using the SGDM optimizer for the aforementioned deep neural networks, respectively. We also developed an Android and IOS mobile application using the novel deep neural network and a flutter framework. The mobile application was tested using brain tumour images collected from SRM hospital and an accuracy of 82.5% was obtained. The untrained data of 155 images collected from hospital were used to validate this application. In terms of execution speed, the proposed architecture surpassed the existing pre-trained models. Thus, when compared to approximately similar approaches, our newly developed network outperformed them.