Novel Volumetric Sub-region Segmentation in Brain Tumors
Subhashis Banerjee1,2, Sushmita Mitra1
1Machine Intelligence Unit, Indian Statistical Institute, Kolkata, India.
Frontiers in Computational Neuroscience
|February 11, 2020
Summary
A novel deep learning model, Multi-Planar Spatial Convolutional Neural Network (MPS-CNN), accurately segments brain tumor sub-regions from MRI scans. This automated approach shows high performance in identifying enhancing tumor and tumor core regions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Accurate segmentation of brain tumors and their sub-regions from multimodal MRI is crucial for diagnosis and treatment planning.
- Existing methods often struggle with precise delineation of complex tumor structures like peritumoral edema and necrotic core.
Purpose of the Study:
- To develop and evaluate a novel deep learning model, the Multi-Planar Spatial Convolutional Neural Network (MPS-CNN), for automated brain tumor segmentation.
- To improve the accuracy of segmenting various tumor sub-regions, including peritumoral edema (ED), necrotic core (NCR), and enhancing (ET) and non-enhancing tumor core (NET).
Main Methods:
- An encoder-decoder CNN architecture was designed for pixel-wise segmentation across axial, sagittal, and coronal planes.
- A consensus fusion strategy combined multi-planar segmentations, refined by a Conditional Random Field (CRF) for post-processing.
- Spatial-pooling and unpooling layers were incorporated to preserve spatial information and reduce boundary segmentation errors.
- An aggregated loss function was developed to address data imbalance issues.
Main Results:
- The MPS-CNN achieved high Dice scores on the BraTS 2018 validation dataset: 0.90216 for whole tumor (WT), 0.87247 for tumor core (TC), and 0.82445 for enhancing tumor (ET).
- The model demonstrated top performance for ET and TC segmentation, excelling in both Dice and Hausdorff distance metrics.
- For WT segmentation, MPS-CNN achieved the second-highest accuracy, closely matching the best-performing method.
Conclusions:
- The proposed MPS-CNN model offers effective and automated segmentation of brain tumor sub-regions from multimodal MRI.
- Its multi-planar approach and CRF-based refinement contribute to high accuracy in delineating tumor boundaries and components.
- MPS-CNN shows significant potential for clinical application in neuro-oncology, outperforming existing methods in key segmentation tasks.
