A triplanar ensemble model for brain tumor segmentation with volumetric multiparametric magnetic resonance images

Abstract

Automated segmentation methods can produce faster segmentation of tumors in medical images, aiding medical professionals in diagnosis and treatment plans. A 3D U-Net method excels in this task but has high computational costs due to large model parameters, which limits their application under resource constraints. This study targets an optimized triplanar (2.5D) model ensemble to generate accurate segmentation with fewer parameters. The proposed triplanar model uses spatial and channel attention mechanisms and information from multiple orthogonal planar views to predict segmentation labels. In particular, we studied the optimum filter size to improve the accuracy without increasing the network complexity. The model generated output is further post-processed to fine-tune the segmentation results. The Dice similarity coefficients (Dice-score) of the Brain Tumor Segmentation (BraTS) 2020 training set for enhancing tumor (ET), whole tumor (WT), and tumor core (TC) are 0.736, 0.896, and 0.841, whereas, for the validation set, they are 0.713, 0.873, and 0.778, respectively. The proposed base model has only 10.25M parameters, three times less than BraTS 2020’s best-performing model (ET 0.798, WT 0.912, TC 0.857) on the validation set. The proposed ensemble model has 93.5M parameters, 1.6 times less than the top-ranked model and two times less than the third-ranked model (ET 0.793, WT 0.911, TC 0.853 on validation set) of BraTS2020 challenge.

Keywords

• Automated segmentation
• Brain tumor
• Ensemble
• Machine learning
• Medical image analysis