Automatic Meningioma Segmentation and Grading Prediction: A Hybrid Deep-Learning Method

Chaoyue Chen, Yisong Cheng, Jianfeng Xu, Ting Zhang, Xin Shu, Wei Huang, Yu Hua, Yang Zhang, Yuen Teng, Lei Zhang, Jianguo XuJournal of Personalized Medicine 2021 · DOI

TL;DR

This paper introduces a novel hybrid deep learning system that automates the segmentation of meningiomas from MRI scans, reconstructs their 3D structure along with surrounding blood vessels, and predicts their grade (benign vs. malignant). This integrated approach aims to improve surgical planning and patient outcome prediction by providing clinicians with a more comprehensive and objective assessment tool.

What Problem Does This Paper Address?

Meningiomas are common intracranial tumors, and their accurate assessment is crucial for effective treatment. Currently, distinguishing between low-grade (benign) and high-grade (malignant) meningiomas, as well as understanding their spatial relationship with critical structures like blood vessels, relies heavily on manual interpretation of MRI scans. This process can be subjective, time-consuming, and prone to inter-observer variability. Furthermore, visualizing the complex 3D anatomy of the tumor and its surrounding vasculature for surgical planning is challenging. This paper addresses the need for an automated, objective, and comprehensive system that can assist clinicians in meningioma diagnosis, grading, and surgical preparation.

Methods

The researchers developed a three-module cascade deep learning system:

  1. Segmentation Module: A modified U-Net convolutional neural network (CNN) was employed for automatic tumor segmentation. Both 2D and 3D U-Net architectures were explored, with the 2D U-Net ultimately showing superior performance in validation.
  2. 3D-Rendering Module: This module takes the segmented tumor images and uses rendering algorithms (isosurface rendering for tumors, Hessian algorithm for vessels) to create a 3D visualization. It reconstructs both the tumor's spatial structure and the intracranial blood vessels, providing a detailed anatomical overview.
  3. Grading Prediction Module: A DenseNet 121 CNN was used to classify meningiomas into low-grade and high-grade categories. This module was found to perform significantly better when trained on segmented tumor images (Plan B) compared to raw images (Plan A).

The system was designed as a cascade, meaning the output of one module serves as the input for the next, creating an end-to-end pipeline.

Datasets

The study utilized two datasets:

  • Primary Dataset: Retrospective MRI scans from 625 patients diagnosed with meningioma at the West China Hospital, Sichuan University, collected between 2008 and 2018. This dataset was used for training and internal validation.
  • External Validation Dataset: MRI scans from 18 low-grade meningioma patients from another institution. This dataset, acquired from different MRI platforms (Siemens MAGNETOM Symphony 1.5T and MAGNETOM Skyra 3T), was used to test the system's robustness. Due to data privacy restrictions, detailed MR sequence parameters for this dataset were unavailable.

Key Insights and Ideas

  • Hybrid Approach Synergy: Combining segmentation, 3D rendering, and grading prediction into a single system offers a more holistic assessment than any single module alone. The segmentation module's accuracy directly impacts the quality of both the rendering and grading.
  • Segmentation as a Foundation: The success of the grading prediction module was significantly boosted by using segmented tumor images as input, highlighting the importance of accurate tumor delineation for feature extraction.
  • Vessel Visualization for Surgical Planning: The ability to automatically render intracranial vessels alongside the tumor is a crucial step towards improving pre-operative surgical planning, allowing surgeons to better understand the tumor's relationship with critical vasculature.
  • Robustness Across Platforms: The system demonstrated reasonable performance on an external dataset acquired from different MRI scanners, suggesting its potential for broader clinical applicability.

Key Results

  • Segmentation Performance: The 2D U-Net model achieved a Dice coefficient of 0.920 ± 0.009 and a Jaccard score of 0.851 ± 0.016 on the validation group. The 3D U-Net performed slightly lower with a Dice of 0.873 ± 0.020.
  • Grading Prediction: The DenseNet model, when trained on segmented images (Plan B), achieved an Area Under the Curve (AUC) of 0.918 ± 0.006 and an accuracy of 0.901 ± 0.039 for classifying tumors into low-grade and high-grade meningiomas. Plan A (using raw images) yielded lower performance (AUC 0.837 ± 0.034, accuracy 0.813 ± 0.086).
  • Rendering: The system successfully reconstructed tumor bodies with sophisticated surfaces and clearly defined structures, along with major intracranial arteries.
  • External Validation: The system showed good performance on the external dataset, with segmentation results being highly consistent with ground-truth labels, particularly on the 3.0T scanner.

Limitations and Future Work

  • Segmentation Limitations: The U-Net models struggled to segment tumors in specific challenging cases, such as those adjacent to thickened dura mater, located in the anterior skull base, or exhibiting complex shapes with necrosis.
  • External Dataset Limitations: The external validation dataset was small (18 patients) and only included low-grade meningiomas, limiting the evaluation of the grading prediction module's performance on this dataset.
  • Need for Manual Correction: For challenging segmentation cases, manual correction of predicted labels was still required.
  • Vessel Rendering Improvement: The vessel rendering module needs further refinement to automatically recognize specific arteries and allow for selective display/hiding of individual vessels.
  • Future Directions:
    • Conducting large-scale, multi-center studies to further validate the system.
  • Exploring more novel CNN architectures for improved segmentation and grading.
  • Integrating other intellectual tasks, such as brain cancer classification, into the system.
  • Improving the vessel rendering to enhance its clinical utility for surgical planning.
  • Enhancing pre-surgery communication with patients by using the system's visualizations.

Summary updated

Abstract

The purpose of this study was to determine whether a deep-learning-based assessment system could facilitate preoperative grading of meningioma. This was a retrospective study conducted at two institutions covering 643 patients. The system, designed with a cascade network structure, was developed using deep-learning technology for automatic tumor detection, visual assessment, and grading prediction. Specifically, a modified U-Net convolutional neural network was first established to segment tumor images. Subsequently, the segmentations were introduced into rendering algorithms for spatial reconstruction and another DenseNet convolutional neural network for grading prediction. The trained models were integrated as a system, and the robustness was tested based on its performance on an external dataset from the second institution involving different magnetic resonance imaging platforms. The results showed that the segment model represented a noteworthy performance with dice coefficients of 0.920 ± 0.009 in the validation group. With accurate segmented tumor images, the rendering model delicately reconstructed the tumor body and clearly displayed the important intracranial vessels. The DenseNet model also achieved high accuracy with an area under the curve of 0.918 ± 0.006 and accuracy of 0.901 ± 0.039 when classifying tumors into low-grade and high-grade meningiomas. Moreover, the system exhibited good performance on the external validation dataset.
BibTeX
@article{chen2021automatic,
  author = {Chaoyue Chen and Yisong Cheng and Jianfeng Xu and Ting Zhang and Xin Shu and Wei Huang and Yu Hua and Yang Zhang and Yuen Teng and Lei Zhang and Jianguo Xu and Jianguo Xu and Jianguo Xu},
  title = {{Automatic Meningioma Segmentation and Grading Prediction: A Hybrid Deep-Learning Method}},
  journal = {Journal of Personalized Medicine},
  year = {2021},
  doi = {10.3390/jpm11080786},
}
Figure