A Lightweight CNN Model Using Depthwise Separable Convolutions for Brain Tumour Classification

Abstract

Every year, the number of patients with brain cancers (BCs) or brain tumours (BTs) increases. This trend emphasises the necessity of a computerised system for rapid and accurate detection during the diagnosis of BTs. This paper presents a lightweight deep learning (DL) model based on a convolutional neural network (CNN) for a fast and accurate BC detector. The core component of the BC detector is a depthwise separable convolution (DSConv) on top of the 24-layer CNN architectures. The usage of DSConv with Adam’s optimiser achieves comparable effectiveness to conventional convolutional layers, although using fewer parameters. Additionally, L2 regularisation, dropout, and data augmentation were implemented to mitigate the issues of overfitting. The proposed model was trained and tested using the publicly available dataset consisting of MRI images collected from 233 patients in Nanfang Hospital and General Hospital, with 3063 images in total. In summary, the DSConv-based CNN model demonstrates an average accuracy of 97.50% and has an average inference time of 2.1 milliseconds per classification. It consistently surpasses 96.50% accuracy in the classification of the three types of BTs. These findings indicate that the model is well-suited for accurate BTs classification, particularly for glioma, meningioma, and pituitary tumours from MRI images.