Data augmentation using learned transformations for one-shot medical image segmentation (1902.09383v2)

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.

Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation

The paper presents an advanced methodology for enhancing one-shot segmentation in medical imaging, particularly focused on MRI brain scans. The primary innovation lies in using data augmentation driven by learned transformations, which bypasses the dependency on extensive labeled datasets typically required for training convolutional neural networks.

Methodological Overview

The authors propose a semi-supervised approach, starting with a singular labeled instance and complementing it with several unlabeled scans. The core idea is to synthesize additional labeled data by learning a model of transformations which encompass both spatial deformations and intensity changes. This model is crucial for simulating variations that emulate real-world differences in anatomy and imaging modalities.

Key Components

1. Automated Data Augmentation: The methodology hinges on creating new training examples from a solitary labeled scan. It employs learned spatial and appearance transformation models to simulate realistic variations by sampling from a pool of unlabeled datasets. These transformations capture both anatomical diversity and imaging differences, producing synthetic data that is more reflective of true clinical variability.
2. Transformation Models:
   • Spatial Transformations are built using a deformation field model, optimized through a learning-based registration approach. The deformation field helps in establishing accurate correspondences between anatomical structures in different scans.
   • Appearance Transformations manage variations in imaging intensity, ensuring that intensity differences do not compromise the anatomical consistency of the synthetic images.

Performance and Evaluation

The proposed method demonstrates substantial improvements over existing state-of-the-art techniques for one-shot segmentation. Notably, when benchmarked against traditional single-atlas segmentation and hand-tuned augmentation strategies, the proposed method achieved superior accuracy across all experiments. The method yielded a mean Dice score increase of 0.045 over the traditional single-atlas approach, underscoring its efficacy.

Implications and Future Directions

The approach's ability to generate robust synthetic data has significant implications for medical imaging. By reducing the reliance on large labeled datasets, it opens pathways for efficient model training in resource-constrained environments, such as real-time clinical settings. Furthermore, the adaptability of this method across different anatomical structures and imaging types suggests broad applicability.

Looking ahead, potential enhancements include exploring richer interpolation techniques between learned transformations and expanding applications beyond brain MRI to other imaging types such as CT scans. The modular nature of the framework also provides avenues for integrating diffeomorphic registration techniques, which could enhance the precision of spatial transformations.

Overall, this work offers a promising step toward resource-efficient medical image analysis, providing a toolset capable of leveraging minimal labeled data to achieve high segmentation accuracy.