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

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.

Fully Automated MRI Segmentation and Volumetric Measurement of Intracranial Meningioma Using Deep Learning

“…The performance of the nnU‐Nets with GMM with the validation sets in this study was similar to or slightly higher than that described in previous studies using U‐Net or its derivatives, but the tumors in our validation sets were much smaller in volume than in previous studies and included 30% of skull base meningiomas, whose locations are known to be difficult to autosegment 17–20,32 . Specifically, skull base meningiomas tend to be overestimated in volume due to adjacent bony structures prone to be mistaken for parts of the tumor.…”

“…Thanks to the advantages of nnU‐Net in analyzing medical images composed of heterogeneous datasets through self‐configuration, we were able to train the models using a training set that partially (17%) consisted of external images without producing noticeable performance degradation as a result of the heterogeneity 12 . This training with a heterogeneous dataset is likely why the performance with the EVS does not fall significantly than with the IVS, unlike what has been observed in previous studies 18,20 . Therefore, we expect that the present model could be greatly utilized in real medical environments where heterogeneous MR device performances and differences in protocols for doctors, hospitals, regions, and countries are present.…”

“…Such models have been developed by extracting features from MR images through a convolutional neural network and then applying supervised learning to predict tumor grade, genotype, and prognosis. However, existing deep learning models only distinguish malignant meningioma from nonmalignant meningioma or differentiate meningioma from tumors such as hemangiopericytoma that have a similar appearance on MR images, and studies focusing on the automated measurement of meningioma volume are rare 18,20,26–31 . If a minute change in tumor volume can be sensitively detected and aggressiveness can be predicted through a deep learning model during the meningioma surveillance period, it could provide useful information for deciding the optimal timing for surgical or radiosurgical intervention.…”

“…Such artificial intelligence (AI)‐based image analysis methods have been applied extensively in the field of neuro‐oncology in recent years, particularly in glioma, through the use of radiomics 13–15 . However, the few AI‐based studies for meningioma that have been published indicate that the corresponding models are insufficient for clinical use 16–21 . Here, we aimed to develop a fully automated segmentation and volumetric measurement tool for meningioma that can be readily applied to clinical practice and validate it.…”

Meningioma Radiomics: At the Nexus of Imaging, Pathology and Biomolecular Characterization

“…In detail, average Dice coefficients were 0.81 ± 0.10 (range: 0.46–0.93) for the total tumor volume, using FLAIR- and contrast enhanced T1-weighted images. Similarly, Chen et al [ 34 ] developed a modified U-Net convolutional neural network segmentation model based on contrast enhanced T1-weighted images, reporting Dice scores of 0.920 ± 0.009 ( Figure 1 ).…”