Domain and Content Adaptive Convolution for Domain Generalization in Medical Image Segmentation

Shishuai Hu, Zehui Liao, Jianpeng Zhang, Yong XiaarXiv 2021 · OpenReview

TL;DR

This paper introduces DCAC, a novel domain generalization method for medical image segmentation that tackles the challenge of varying image quality across different sources. By employing "domain adaptive convolutions" (DAC) and "content adaptive convolutions" (CAC), DCAC dynamically adjusts its parameters to better segment unseen medical images, outperforming existing methods on prostate, COVID-19 lesion, and optic cup/disc segmentation tasks.

What Problem Does This Paper Address?

Medical image segmentation models trained in a controlled lab environment often struggle when applied to real-world clinical data. This performance drop is primarily due to the "domain gap" – variations in image quality stemming from different scanners, imaging protocols, and patient populations. Traditional domain generalization methods often rely on static convolutions, limiting their ability to adapt to these unseen variations. This paper aims to bridge this gap by developing a more flexible and adaptive segmentation model that can generalize effectively to diverse medical imaging domains.

Methods

The core innovation of this work is the Domain and Content Adaptive Convolution (DCAC) model. It builds upon a standard encoder-decoder architecture for segmentation. The key components are:

  • Domain Adaptive Convolution (DAC) Module: This module uses a "domain predictor" to estimate a "domain code" for each input image, indicating its similarity to known source domains. A "domain-aware controller" then uses this code to dynamically generate parameters for a dynamic convolutional layer. This allows the model to adapt to the specific characteristics of the unseen target domain.
  • Content Adaptive Convolution (CAC) Module: This module employs a "content-aware controller" that takes global image features as input. It dynamically generates parameters for a series of dynamic convolutional layers. This allows the model to adapt to the specific content of each individual test image, further enhancing its robustness.

These adaptive modules are integrated into the encoder-decoder backbone, allowing for dynamic adjustments at multiple levels. The model is trained using a combination of segmentation and domain classification losses, with a gradient truncation strategy to balance these objectives.

Datasets

The proposed DCAC model was evaluated on three distinct medical image segmentation tasks using publicly available datasets:

  • Prostate Segmentation: MRI scans from six different domains.
  • COVID-19 Lesion Segmentation: CT scans from four different domains.
  • Optic Cup/Optic Disc (OC/OD) Segmentation: Fundus images from four different public datasets, known for their inconsistent statistical characteristics.

Key Insights and Ideas

  • Dynamic Adaptation is Crucial: The paper strongly suggests that static models are inherently limited in handling the diverse domain shifts common in medical imaging. Dynamic convolutions, which adapt their parameters based on input, are key to achieving robust generalization.
  • Two-Pronged Adaptation: The combination of domain-specific adaptation (DAC) and content-specific adaptation (CAC) is more effective than using either alone. This dual approach allows the model to address both broad domain variations and fine-grained image-level differences.
  • Domain Code as a Bridge: The "domain code" effectively captures the relationship between source and target domains, enabling the DAC module to steer the model's adaptation process.
  • Multi-scale Features Enhance Domain Prediction: Using multi-scale features from the encoder for domain prediction leads to more accurate domain classification, which in turn benefits the overall segmentation performance.

Key Results

The DCAC model demonstrated superior performance across all three segmentation tasks compared to baseline methods and four state-of-the-art domain generalization techniques.

  • Prostate Segmentation: DCAC achieved the highest average Dice Similarity Coefficient (DSC) and lowest Average Surface Distance (ASD) across all domains. For instance, on Domain 2, DCAC achieved a DSC of 90.51% with an ASD of 0.89, outperforming the best competing method.
  • COVID-19 Lesion Segmentation: DCAC also showed improved performance, with an average DSC of 65.10% and ASD of 12.76.
  • OC/OD Segmentation: DCAC achieved the highest average DSC (88.47%) and lowest ASD (11.32%) on this challenging task.
  • Ablation Studies: Experiments confirmed the effectiveness of both DAC and CAC modules, with their combined use yielding the best results. Variants using only one module or a unified dynamic head performed worse.

Limitations and Future Work

  • Multi-Source Requirement: The current DCAC model is designed for multi-source domain generalization and requires data from multiple source domains for training. It cannot be directly applied to single-source domain generalization scenarios.
  • Performance Discrepancy: The performance gains of DCAC on CT images were less pronounced compared to MR and fundus images. The authors attribute this to the generally more consistent image quality of CT scans, whereas MR and fundus images exhibit more vendor-specific variations.

The authors suggest extending the proposed DCAC model to address these limitations in future work.

Summary updated

Abstract

The domain gap caused mainly by variable medical image quality renders a major obstacle on the path between training a segmentation model in the lab and applying the trained model to unseen clinical data. To address this issue, domain generalization methods have been proposed, which however usually use static convolutions and are less flexible. In this paper, we propose a multi-source domain generalization model based on the domain and content adaptive convolution (DCAC) for the segmentation of medical images across different modalities. Specifically, we design the domain adaptive convolution (DAC) module and content adaptive convolution (CAC) module and incorporate both into an encoder-decoder backbone. In the DAC module, a dynamic convolutional head is conditioned on the predicted domain code of the input to make our model adapt to the unseen target domain. In the CAC module, a dynamic convolutional head is conditioned on the global image features to make our model adapt to the test image. We evaluated the DCAC model against the baseline and four state-of-the-art domain generalization methods on the prostate segmentation, COVID-19 lesion segmentation, and optic cup/optic disc segmentation tasks. Our results not only indicate that the proposed DCAC model outperforms all competing methods on each segmentation task but also demonstrate the effectiveness of the DAC and CAC modules. Code is available at \url{https://git.io/DCAC}.
BibTeX
@misc{hu2021domain,
  author = {Shishuai Hu and Zehui Liao and Jianpeng Zhang and Yong Xia},
  title = {{Domain and Content Adaptive Convolution for Domain Generalization in Medical Image Segmentation}},
  year = {2021},
  eprint = {2109.05676},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.05676},
}
Figure