Abstract
Automatic segmentation of vestibular schwannoma (VS) tu-
mors from magnetic resonance imaging (MRI) would facilitate efficient
and accurate volume measurement to guide patient management and
improve clinical workflow. The accuracy and robustness is challenged by
low contrast, small target region and low through-plane resolution. We
introduce a 2.5D convolutional neural network (CNN) able to exploit the
different in-plane and through-plane resolutions encountered in standard
of care imaging protocols. We propose an attention module with explicit
supervision on the attention maps to enable the CNN to focus on the
small target for more accurate segmentation. We also propose a hardness-
weighted Dice loss function that gives higher weights to harder voxels to
boost the training of CNNs. Experiments with ablation studies on the
VS tumor segmentation task show that: 1) our 2.5D CNN outperforms
its 2D and 3D counterparts, 2) our supervised attention mechanism out-
performs unsupervised attention, 3) the voxel-level hardness-weighted
Dice loss improves the segmentation accuracy. Our method achieved an
average Dice score and ASSD of 0.87 and 0.43 mm respectively. This will
facilitate patient management decisions in clinical practice.
mors from magnetic resonance imaging (MRI) would facilitate efficient
and accurate volume measurement to guide patient management and
improve clinical workflow. The accuracy and robustness is challenged by
low contrast, small target region and low through-plane resolution. We
introduce a 2.5D convolutional neural network (CNN) able to exploit the
different in-plane and through-plane resolutions encountered in standard
of care imaging protocols. We propose an attention module with explicit
supervision on the attention maps to enable the CNN to focus on the
small target for more accurate segmentation. We also propose a hardness-
weighted Dice loss function that gives higher weights to harder voxels to
boost the training of CNNs. Experiments with ablation studies on the
VS tumor segmentation task show that: 1) our 2.5D CNN outperforms
its 2D and 3D counterparts, 2) our supervised attention mechanism out-
performs unsupervised attention, 3) the voxel-level hardness-weighted
Dice loss improves the segmentation accuracy. Our method achieved an
average Dice score and ASSD of 0.87 and 0.43 mm respectively. This will
facilitate patient management decisions in clinical practice.
| Original language | English |
|---|---|
| Title of host publication | Medical Image Computing and Computer Assisted Intervention – MICCAI 2019 |
| Publication status | Accepted/In press - 5 Jun 2019 |