Early Diagnosis of Brain Cancer with Deep Learning
Frank Hui, Brian Ji
Burnaby North Secondary
Floor Location : S 012 H
Despite great advances in the field of oncology, glioblastoma multiforme's extreme aggression still results in a grim prognosis. A median treated survival length of a mere 11-15 months, couples with one of the lowest survival rates of all cancers, at ~4%. Furthermore, glioblastoma multiforme's lethality is compounded by its relative prevalence: Glioblastoma multiforme (GBM) accounts for 80% of malignant gliomas and is one of the most common and deadly brain malignancies, with approximately 13200 cases per annum within the United States alone.
Simultaneously, detection models have made great progress with non-small cell lung cancer and Alzheimer's in the medical diagnostics industry, while the current conventional detection of brain tumors involves human inspection of radiological imagery for tissue abnormalities. Our project aims to utilize convolutional neural networks on MR imaging for the early detection of GBM. Work was primarily split into three categories: data extraction and preprocessing of imagery, convolutional neural network training, and the machine learning classification period.
Our results were promising, displaying a 98% testing accuracy (30018 test-set) after 10 epochs of training (270179 training-set) with 1239 patients. The automation of early-stage tumor detection drastically reduces the workload of radiologists, aids with patient outcomes through earlier treatment, and may provide insight into the characteristics of high-grade astrocytomas. MR Imaging and ML algorithms look promising regarding their potential applications in the medical field, particularly in the field of medical diagnoses.