Automatic Segmentation of Meningioma on 3D MRI Using Convolutional Neural Networks

Overview

As meningioma accounts for 40% of adult brain tumors worldwide, timely tumor assessment is required to ensure patient health. Currently, manual tumor segmentation imposes a high burden on the radiologists and lacks intra- and inter-observer repeatability. Furthermore, many of the current methods for tumor assessment in neuro-oncology practice are based on 2D tumor measurements which cannot provide accurate information about the tumor and its response to treatments. Algorithmic approaches can also be subject to inefficiencies and bias. This project develops an automated segmentation framework that can augment tumor assessments and improve prediction of patient prognosis by producing reliable and reproducible volumetric measurements. Four types of magnetic resonance images (MRIs), including T1, T1-CE, T2, and T2-FLAIR scans, were used from a collection of 358 meningioma patients. The dataset was split into 80% for training, 15% for validation, and 5% for testing. An automated segmentation model was trained based on the MRI scans using the nnU-Net architecture. The nnU-Net model classified regions of the tumor on MRI scans into edema (ED) or enhancing tumor (ET) regions. The prediction of the nnU-Net model was cross-checked with the ground truth labels to assess prediction accuracy. The model inference was run on both the validation and test datasets. The model achieved a mean DICE score of 0.88 and mean sensitivity of 86.3% for segmentation of ET, and a mean DICE score of 0.88 and mean sensitivity of 90.1% for segmentation of ED. The trained model achieved a mean DICE score of 0.89 and mean sensitivity of 88.3%. The automated segmentation method created in this project based on nnU-Net framework shows reliable performance for tumor subregion segmentation of meningioma.

Competition history

• AJAS 2024 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• AJAS None projects
• Every AJAS 2024 project