[Paper Review] Data augmentation using learned transformations for one-shot medical image segmentation

Amy Zhao, Guha Balakrishnan+3 more|arXiv (Cornell University)|Feb 25, 2019

Medical Image Segmentation Techniques79 references57 citations

TL;DR

The paper presents a semi-supervised data augmentation method that learns spatial and appearance transformations from unlabeled MRI scans to synthesize labeled examples from a single atlas, improving one-shot brain MRI segmentation. It outperforms hand-tuned augmentation and single-atlas baselines, approaching fully supervised performance.

ABSTRACT

Image segmentation is an important task in many medical applications. Methods based on convolutional neural networks attain state-of-the-art accuracy; however, they typically rely on supervised training with large labeled datasets. Labeling medical images requires significant expertise and time, and typical hand-tuned approaches for data augmentation fail to capture the complex variations in such images. We present an automated data augmentation method for synthesizing labeled medical images. We demonstrate our method on the task of segmenting magnetic resonance imaging (MRI) brain scans. Our method requires only a single segmented scan, and leverages other unlabeled scans in a semi-supervised approach. We learn a model of transformations from the images, and use the model along with the labeled example to synthesize additional labeled examples. Each transformation is comprised of a spatial deformation field and an intensity change, enabling the synthesis of complex effects such as variations in anatomy and image acquisition procedures. We show that training a supervised segmenter with these new examples provides significant improvements over state-of-the-art methods for one-shot biomedical image segmentation. Our code is available at https://github.com/xamyzhao/brainstorm.

Motivation & Objective

• Address the challenge of limited labeled data in medical image segmentation.
• Develop an automated augmentation framework that leverages unlabeled scans.
• Synthesize diverse labeled training examples by modeling spatial and appearance transformations.
• Demonstrate improvements over state-of-the-art one-shot segmentation methods on brain MRI data.

Proposed method

• Model a spatial transformation using a learned displacement field via a VoxelMorph-based registration network.
• Model an appearance/intensity transformation as a per-voxel addition learned with a semantically-aware smoothness regularizer.
• Synthesize labeled examples by applying independent spatial and appearance transforms to a single labeled atlas, ensuring labels follow the warped image.
• Train a supervised segmentation network on the augmented labeled dataset.
• Evaluate using Dice scores across 30 brain structures on a held-out test set.
• Implementation uses differentiable 3D spatial transformer layers and trains transforms with a dedicated loss combining image similarity and smoothness terms.

Experimental results

Research questions

• RQ1 Can learned spatial and appearance transformations from unlabeled MRIs produce realistic labeled augmentations when applied to a single atlas?
• RQ2 Do these learned augmentations improve one-shot brain MRI segmentation beyond hand-tuned augmentation and single-atlas methods?
• RQ3 How close can one-shot segmentation get to fully supervised performance using synthesized examples?
• RQ4 Is independence of spatial and appearance transforms advantageous for generating diverse training data?

Key findings

• Ours-indep achieves Dice score of 0.804 (0.130) on average, outperforming SAS baseline and rand-aug.
• Ours-indep + rand-aug attains 0.815 (0.123) Dice on average, the best among one-shot methods.
• Ours-coupled reaches 0.795 (0.133), better than SAS and SAS-aug baselines.
• SAS baseline Dice score is 0.759 (0.137) and SAS-aug is 0.775 (0.147).
• Fully supervised upper bound yields 0.849 (0.092).
• The proposed method yields significant improvements over baselines (p<1e-15 vs rand-aug; p<1e-20 vs SAS) across 100 test subjects.