MRI-Based Meningioma Firmness Classification Using an Adversarial Feature Learning Approach

TL;DR

This paper presents a new machine learning method using MRI scans to classify meningioma firmness, improving accuracy and reducing subjectivity in surgical planning.

Contribution

The novel adversarial feature learning approach combines BiGAN and depth-wise separable models for meningioma firmness classification.

Findings

The proposed model achieved 94.7% accuracy in classifying meningioma firmness.

The BiGAN-based feature extraction outperformed state-of-the-art methods in classification performance.

The model's use of unsupervised learning reduced reliance on manual feature engineering.

Abstract

The firmness of meningiomas is a critical factor that impacts the surgical approach recommended for patients. The conventional approaches that couple image processing techniques with radiologists’ visual assessments of magnetic resonance imaging (MRI) proved to be time-consuming and subjective to the physician’s judgment. Recently, machine learning-based methods have emerged to classify MRI instances into firm or soft categories. Typically, such solutions rely on hand-crafted attributes and/or feature engineering techniques to encode the visual content of patient MRIs. This research introduces a novel adversarial feature learning approach to tackle meningioma firmness classification. Specifically, we present two key contributions: (i) an unsupervised feature extraction approach utilizing the Bidirectional Generative Adversarial Network (BiGAN) and (ii) a depth-wise separable deep learning model were designed to map the relevant MRI features with the predefined meningioma firmness classes. The experiments demonstrated that associating the BiGAN encoder, for unsupervised feature extraction, with a depth-wise separable deep learning model enhances the classification performance. Moreover, the proposed pre-trained BiGAN encoder-based model outperformed relevant state-of-the-art methods in meningioma firmness classification. It achieved an accuracy of 94.7% and a weighted F1-score of 95.0%. This showcases the proposed model’s ability to extract discriminative features and accurately classify meningioma consistency.

Figures