Semi-Supervised Medical Image Segmentation via Learning Consistency under Transformations (1911.01218v1)

Abstract: The scarcity of labeled data often limits the application of supervised deep learning techniques for medical image segmentation. This has motivated the development of semi-supervised techniques that learn from a mixture of labeled and unlabeled images. In this paper, we propose a novel semi-supervised method that, in addition to supervised learning on labeled training images, learns to predict segmentations consistent under a given class of transformations on both labeled and unlabeled images. More specifically, in this work we explore learning equivariance to elastic deformations. We implement this through: 1) a Siamese architecture with two identical branches, each of which receives a differently transformed image, and 2) a composite loss function with a supervised segmentation loss term and an unsupervised term that encourages segmentation consistency between the predictions of the two branches. We evaluate the method on a public dataset of chest radiographs with segmentations of anatomical structures using 5-fold cross-validation. The proposed method reaches significantly higher segmentation accuracy compared to supervised learning. This is due to learning transformation consistency on both labeled and unlabeled images, with the latter contributing the most. We achieve the performance comparable to state-of-the-art chest X-ray segmentation methods while using substantially fewer labeled images.

• The paper proposes a semi-supervised method that leverages unlabeled data and learns segmentation consistency under transformations to improve medical image segmentation accuracy.
• Evaluated on the JSRT chest X-ray dataset, the method significantly enhances accuracy compared to supervised approaches, particularly with limited labeled data.
• This research demonstrates a practical way to reduce costly data annotation in medical imaging while showcasing the potential of consistency-based learning from unlabeled data.

Overview of Semi-Supervised Medical Image Segmentation via Learning Consistency under Transformations

The paper "Semi-Supervised Medical Image Segmentation via Learning Consistency under Transformations" by Bortsova et al. addresses a pertinent challenge in medical image processing—the scarcity of labeled data required by supervised deep learning algorithms for efficient image segmentation, specifically in the medical domain. By leveraging a semi-supervised approach, the authors propose a novel methodology that not only utilizes labeled data but also effectively incorporates a wealth of unlabeled data to improve segmentation performance.

Methodology

The authors introduce a semi-supervised technique that emphasizes learning segmentation consistency under transformations. The core idea involves employing a Siamese architecture comprising two identical branches. Each branch of the network processes differently transformed versions of an input image. The architecture is subjected to a composite loss function that integrates a supervised segmentation loss with an additional transformation consistency term applied to both labeled and unlabeled images. Specifically, the transformation that the method targets is equivariance to elastic deformations—common in medical imaging applications.

The supervised loss term is applied to labeled images, whereas the unsupervised consistency term aims to ensure that segmentations remain invariant under these transformations for both labeled and unlabeled images. The network, thus leveraging both labeled and unlabeled datasets, aims to reach higher accuracy.

Evaluation and Findings

The method is validated on the JSRT chest X-ray dataset through rigorous 5-fold cross-validation. The results are compelling, showcasing that the proposed semi-supervised technique significantly enhances segmentation accuracy compared to a purely supervised approach. This advantage is distinctly pronounced when the size of the labeled dataset is small. The research highlights that learning transformation consistency, especially from unlabeled data, results in performance metrics comparable to state-of-the-art methods, while considerably reducing the reliance on labeled datasets.

Implications

The implications of this research are profound both in practical and theoretical realms of medical image analysis. Practically, the reduction in required labeled data without compromising on accuracy can substantially lower costs and time associated with data annotation in medical imaging. Theoretically, this research underscores the potential and efficacy of leveraging unlabeled data in deep learning models via consistency-based learning—showcasing a forward-thinking approach towards robust semi-supervised methods.

Future Directions

Future developments could explore the generalizability of this method across various medical imaging modalities that involve diverse transformation types. Expanding beyond elastic deformations, potentially integrating additional transformations like affine or non-linear warps, could further optimize the proposed framework. Furthermore, extending this methodology to three-dimensional medical imaging datasets presents an intriguing avenue for broader applicability.

In summary, the paper by Bortsova et al. demonstrates an insightful blend of semi-supervised learning and transformation-based consistency, presenting a viable pathway for enhancing segmentation accuracy in medical imaging while alleviating the burden of extensive labeled data requirements.