Deep Learning-Based Brain Tumor Diagnosis on Smartphones Using Optimized MobileNetV2 Models

Abstract

Identifying brain tumors early and accurately is a significant way to improve patient outcomes, but access to numerous advanced diagnostic tools is not standardized across the world due to the cost and availability of MRI scans. We present a lightweight smartphone brain tumor diagnostic tool with deep learning–based diagnostic decision support in a contextualized way. We developed a convolutional neural network (CNN) based on MobileNetV2 for mobile deployment that allows for the processing of MRI images on consumer smartphones in real-time. The model was developed and validated on a publicly available brain tumor MRI dataset of glioma, meningioma, pituitary tumor, and normal cases, achieving an overall accuracy of 98% and classifying cases in less than 100 ms on standard iOS devices. This work demonstrates that with a lightweight architecture and on-device processing for the medical image, diagnostic decision support can be facilitated in a cost-effective, portable way, while also creating confidence factors in patient privacy, and represents an immense opportunity in lower-resourced clinical care, telemedicine, and point-of-care diagnosis around patients. It demonstrates another methodological option for the feasible implementation of advanced deep learning models to assist significant medical imaging workflows in a smartphone device.