Detection and Grading of Gliomas Using a Novel Two-Phase Machine Learning Method Based on MRI Images

Tao Chen, Feng Xiao, Zunpeng Yu, Mengxue Yuan, Haibo Xu, Long LuFrontiers in Neuroscience 2021 · DOI

Abstract

The early detection and grading of gliomas is important for treatment decision and assessment of prognosis. Over the last decade numerous automated computer analysis tools have been proposed, which can potentially lead to more reliable and reproducible brain tumor diagnostic procedures. In this paper, we used the gradient-based features extracted from structural magnetic resonance imaging (sMRI) images to depict the subtle changes within brains of patients with gliomas. Based on the gradient features, we proposed a novel two-phase classification framework for detection and grading of gliomas. In the first phase, the probability of each local feature being related to different types (e.g., diseased or healthy for detection, benign or malignant for grading) was calculated. Then the high-level feature representing the whole MRI image was generated by concatenating the membership probability of each local feature. In the second phase, the supervised classification algorithm was used to train a classifier based on the high-level features and patient labels of the training subjects. We applied this framework on the brain imaging data collected from Zhongnan Hospital of Wuhan University for glioma detection, and the public TCIA datasets including glioblastomas (WHO IV) and low-grade gliomas (WHO II and III) data for glioma grading. The experimental results showed that the gradient-based classification framework could be a promising tool for automatic diagnosis of brain tumors.

Introduction

Gliomas are the most common primary brain tumors, and their "grade" (from WHO Grade II to IV) dictates whether a patient receives conservative observation or aggressive chemotherapy and radiation. Currently, diagnosis relies heavily on visual inspection by neuroradiologists or manual "Region of Interest" (ROI) segmentation—both of which are labor-intensive, subjective, and prone to error.

While deep learning has shown promise in this area, it often acts as a "black box" and requires massive amounts of data. Furthermore, 3D MRI data is high-dimensional; simply feeding raw pixels into a classifier leads to the "curse of dimensionality," where the model overfits the noise rather than learning the actual pathology. This paper seeks an automated, interpretable, and computationally efficient way to identify and grade tumors across the entire volume of the brain using a novel two-phase framework.

Methodology: The Two-Phase Framework

The researchers developed a "local-to-whole" approach that transforms raw image textures into semantic probability maps before final classification.

Phase 1: Feature Extraction and Semantic Mapping

The team utilized Histogram of Oriented Gradients (HOG), a technique traditionally used in computer vision for pedestrian detection. HOG captures the distribution of intensity gradients, making it excellent at depicting the subtle "occupying effects" and structural changes a tumor causes in brain tissue.

The process of extracting Histogram of Oriented Gradients (HOG) features from MRI slices, involving grid division (A), gradient calculation (B), orientation binning (C), and histogram generation (D).

Instead of flattening all HOG features into one giant vector, the researchers analyzed each local "cell" of the MRI independently. They used Fuzzy C-means clustering to assign each cell a "membership probability" (e.g., asking how likely a specific small patch of tissue is to be diseased vs. healthy). This reduces the high-dimensional HOG data into a compact, semantic map. By using fuzzy logic instead of hard clustering, the model better accounts for the ambiguity in transition zones between healthy tissue and malignant cells.

Phase 2: Global Classification

These local probabilities were concatenated to represent the entire brain and fed into a Support Vector Machine (SVM). The SVM then makes the final diagnostic call: detecting the presence of a tumor or determining its grade (High-Grade/GBM vs. Low-Grade/LGG).

Experimental Setup

The study utilized two distinct datasets focusing on T2-FLAIR MRI sequences, known for high contrast in tumorous tissue:

  • DS-Detect: 99 subjects (62 with glioma, 37 healthy controls) from Zhongnan Hospital of Wuhan University.
  • DS-Grade: 134 subjects (76 Glioblastoma/Grade IV and 58 Low-Grade Glioma/Grade II & III) from the public TCIA (The Cancer Imaging Archive).

Key Results

The two-phase method significantly outperformed baseline approaches that simply concatenated raw HOG features:

  • Glioma Detection: Achieved 86.3% accuracy and an AUC of 0.921. The model was particularly strong at detection, with a sensitivity of 89.4%.
  • Glioma Grading: Achieved 76.3% accuracy and an AUC of 0.806.

The authors found that detection is significantly easier than grading. Tumors cause obvious structural displacements that are easy for HOG to catch, whereas grading requires distinguishing between much subtler morphological differences within the tumor itself.

Discussion and Future Work

By converting complex texture descriptors into a single probability value per cell, the authors effectively "compressed" the data, allowing the SVM to learn from meaningful patterns rather than pixel-level noise. Unlike many models that require a human to select the "best" slice, this pipeline processes the entire 3D volume automatically.

However, limitations remain. The study relied only on T2-FLAIR images; incorporating other sequences like T1-weighted contrast-enhanced or Diffusion-Weighted Imaging could provide a more holistic view of tumor necrosis. Additionally, the "Low-Grade Glioma" category was a mix of different histological types, which likely made the grading task more difficult. Future work will focus on integrating clinical factors like patient age and developing direct interfaces with hospital PACS systems for real-time automated "second opinions."

Summary updated

Abstract

The early detection and grading of gliomas is important for treatment decision and assessment of prognosis. Over the last decade numerous automated computer analysis tools have been proposed, which can potentially lead to more reliable and reproducible brain tumor diagnostic procedures. In this paper, we used the gradient-based features extracted from structural magnetic resonance imaging (sMRI) images to depict the subtle changes within brains of patients with gliomas. Based on the gradient features, we proposed a novel two-phase classification framework for detection and grading of gliomas. In the first phase, the probability of each local feature being related to different types (e.g., diseased or healthy for detection, benign or malignant for grading) was calculated. Then the high-level feature representing the whole MRI image was generated by concatenating the membership probability of each local feature. In the second phase, the supervised classification algorithm was used to train a classifier based on the high-level features and patient labels of the training subjects. We applied this framework on the brain imaging data collected from Zhongnan Hospital of Wuhan University for glioma detection, and the public TCIA datasets including glioblastomas (WHO IV) and low-grade gliomas (WHO II and III) data for glioma grading. The experimental results showed that the gradient-based classification framework could be a promising tool for automatic diagnosis of brain tumors.
BibTeX
@inproceedings{chen2021detection,
  author = {Tao Chen and Feng Xiao and Zunpeng Yu and Mengxue Yuan and Haibo Xu and Long Lu},
  title = {{Detection and Grading of Gliomas Using a Novel Two-Phase Machine Learning Method Based on MRI Images}},
  booktitle = {Frontiers in Neuroscience},
  year = {2021},
  doi = {10.3389/fnins.2021.650629},
}
Figure