# Adult-type diffuse glioma prediction using MnasNet optimized by the advanced single candidate optimizer

## TL;DR

This paper presents a non-invasive deep learning method using MRI scans to accurately predict adult-type diffuse glioma, outperforming existing techniques.

## Contribution

A novel deep learning model using MnasNet optimized by the Advanced Single Candidate Optimizer with enhanced learning techniques for glioma prediction.

## Key findings

- The proposed method achieved high sensitivity (95.11%) and specificity (96.57%) in predicting adult-type diffuse glioma.
- It outperformed six state-of-the-art techniques in accuracy and other evaluation metrics.
- The model was validated on 533 patients with strict 10-fold cross-validation.

## Abstract

Diffuse glioma is the most common and aggressive type of the primary brain tumor of adults that has few treatment options with poor prognosis. Existing diagnostic methods including biopsy and histopathological examination are invasive, time consuming and prone to inter-observer variations. To overcome these shortcomings, this paper suggests a non-invasive, deep learning-based approach to the prediction of adult-type diffuse glioma using preoperative T2-weighted MRI. The paradigm incorporates the alternated MnasNet design which is optimized by a new metaheuristic-based algorithm called the Advanced Single Candidate Optimizer (ASCO) but with the addition of opposition-based learning and Chebyshev chaotic map. The method was trained and tested on a pooled set of 533 patients, 237 of Nagoya University Hospital and 296 of a publicly accessible database coordinated of ground-truth (IDH mutation and 1p/19q codeletion status). Strict 10-fold cross-validation was conducted on an independent test set with a sensitivity of 95.11, specificity of 96.57, precision of 98.75, accuracy of 97.30, F1-score of 97.76 and a Matthews Correlation Coefficient of 92.62. Comparative analyses shows the best result toward six state of the art techniques to prove the clinical potential of the proposed system to predict glioma accurately and non-invasively.