Transformation Consistent Self-ensembling Model for Semi-supervised Medical Image Segmentation (1903.00348v3)

Abstract: Deep convolutional neural networks have achieved remarkable progress on a variety of medical image computing tasks. A common problem when applying supervised deep learning methods to medical images is the lack of labeled data, which is very expensive and time-consuming to be collected. In this paper, we present a novel semi-supervised method for medical image segmentation, where the network is optimized by the weighted combination of a common supervised loss for labeled inputs only and a regularization loss for both labeled and unlabeled data. To utilize the unlabeled data, our method encourages the consistent predictions of the network-in-training for the same input under different regularizations. Aiming for the semi-supervised segmentation problem, we enhance the effect of regularization for pixel-level predictions by introducing a transformation, including rotation and flipping, consistent scheme in our self-ensembling model. With the aim of semi-supervised segmentation tasks, we introduce a transformation consistent strategy in our self-ensembling model to enhance the regularization effect for pixel-level predictions. We have extensively validated the proposed semi-supervised method on three typical yet challenging medical image segmentation tasks: (i) skin lesion segmentation from dermoscopy images on International Skin Imaging Collaboration (ISIC) 2017 dataset, (ii) optic disc segmentation from fundus images on Retinal Fundus Glaucoma Challenge (REFUGE) dataset, and (iii) liver segmentation from volumetric CT scans on Liver Tumor Segmentation Challenge (LiTS) dataset. Compared to the state-of-the-arts, our proposed method shows superior segmentation performance on challenging 2D/3D medical images, demonstrating the effectiveness of our semi-supervised method for medical image segmentation.

Overview of "Transformation-consistent Self-ensembling Model for Semi-supervised Medical Image Segmentation"

This paper introduces a novel approach for semi-supervised medical image segmentation, addressing a critical limitation faced by supervised deep learning—the dependency on extensive labeled datasets. The proposed model, termed the Transformation-consistent Self-ensembling Model (TCSM_v2), capitalizes on abundant unlabeled data to enhance segmentation accuracy in medical imaging tasks, specifically in conditions where acquiring labeled data is resource-intensive.

Methodology

The technique builds upon self-ensembling models, which have shown success in semi-supervised classification but extends their application to segmentation tasks. The primary innovation in TCSM_v2 is the introduction of a transformation-consistent strategy that leverages less-supervised data to improve model predictions. This is achieved by encouraging the model to produce consistent results for the same input subjected to different transformations. This approach is implemented in a teacher-student framework, where the teacher model is an exponential moving average of the student model, thereby constructing more reliable targets for training.

Key transformations include rotations, flips, and scaling, which are applied to the input images. The unsupervised component of the training process minimizes mean square error loss between transformed outputs, thereby regularizing predictions to be consistent across various perturbations. The loss function combines a supervised cross-entropy term with an unsupervised consistency term, effectively balancing between labeled and unlabeled data utilization.

Experimental Results

The authors demonstrate the efficacy of TCSM_v2 across three medical image segmentation tasks: skin lesions from dermoscopy images, optic disc from retinal fundus images, and liver from CT scans. The method consistently outperforms existing semi-supervised segmentations and state-of-the-art approaches on these datasets. Specifically, notable improvements are reported in terms of Jaccard Index (JA), Dice Coefficient (DI), and accuracy metrics, surpassing several competitive methods optimized for fully supervised settings.

For instance, in dermoscopy image segmentation, the model trained with a mere 50 labeled images and 1950 unlabeled images achieved a JA improvement of over 4% compared to a supervised baseline. Similarly, significant enhancements were observed in retinal fundus and liver CT scan datasets, particularly notable under settings with limited labeled data.

Practical Implications and Theoretical Considerations

The proposed framework is adaptable and demonstrates robust performance improvements by effectively harnessing unlabeled data. For practical applications in clinical settings, this approach offers a promising avenue to reduce the burden of manual annotation while maintaining, or even enhancing, segmentation accuracy. This capability is crucial for integrating AI systems in healthcare, where obtaining large labeled datasets is often infeasible.

Theoretically, TCSM_v2 exploits the natural equivariance properties of transformations in images to enforce regularization, a principle that could extend beyond medical imaging and prove beneficial in other domains requiring high-precision segmentation. Future work may explore further automation in optimizing transformation strategies or integrating domain adaptation techniques to handle datasets with distribution shifts.

Overall, this research contributes a significant advancement in semi-supervised learning, particularly tailored for medical image segmentation, showcasing the potential of leveraging unlabeled medical data to achieve superior segmentation outcomes with minimal supervision.