- Abstract: Automatic segmentation allows advancement in medical diagnosis and follow-up but remains a challenging task. Thanks to new machine learning approaches, this task tends to be more and more robust but still required many manual segmentations. Here we proposed to improve segmentation results obtained by multi-atlas segmentation with corrective learning (CL) approach using a selection of atlases based on morphological similarity to the image to process. We first introduce our morphological measurement dedicated for quadriceps segmentation of 3D T1 Water-only MR images and then use it to select closest atlases. Our results show that using few atlases (3 in lieu of 6) based on our morphological measurement improves segmentation quality and decrease computational time for multi-atlas segmentation with CL. Based on the measurements, we also defined a data augmentation strategy to train U-Net (a well-known and efficient deep learning segmentation approach), expecting better generalization capability, with very promising results.
- Track: full conference paper
- Keywords: automatic segmentation, machine learning, weakly supervised, magnetic resonance imaging, data augmentation
- Paper Type: both