Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ConvLoRA and AdaBN based Domain Adaptation via Self-Training

Sidra Aleem, Julia Dietlmeier, Eric Arazo, Suzanne Little

TL;DR

This work tackles unsupervised multi-target domain adaptation for medical image segmentation under cross-site shifts. It introduces ConvLoRA, a CNN-specific low-rank adapter that freezes pre-trained weights and learns a compact update $W_{PT ext{CONV}} + XY$ with rank $r$, together with Adaptive Batch Normalization (AdaBN) that uses target-domain statistics. The method is integrated into a UDAS-based pipeline with self-training, training only the ConvLoRA parameters and achieving substantial parameter reductions (≈99% fewer trainable parameters) while delivering competitive or superior segmentation performance across multiple target domains. The results demonstrate strong practical impact for MTDA in medical imaging, enabling efficient adaptation across diverse centers without maintaining separate full-model copies per target domain. Key ideas are applicable to CNN architectures with BN layers and are supported by extensive ablations on adapter placement, BN use, and training duration.

Abstract

Existing domain adaptation (DA) methods often involve pre-training on the source domain and fine-tuning on the target domain. For multi-target domain adaptation, having a dedicated/separate fine-tuned network for each target domain, that retain all the pre-trained model parameters, is prohibitively expensive. To address this limitation, we propose Convolutional Low-Rank Adaptation (ConvLoRA). ConvLoRA freezes pre-trained model weights, adds trainable low-rank decomposition matrices to convolutional layers, and backpropagates the gradient through these matrices thus greatly reducing the number of trainable parameters. To further boost adaptation, we utilize Adaptive Batch Normalization (AdaBN) which computes target-specific running statistics and use it along with ConvLoRA. Our method has fewer trainable parameters and performs better or on-par with large independent fine-tuned networks (with less than 0.9% trainable parameters of the total base model) when tested on the segmentation of Calgary-Campinas dataset containing brain MRI images. Our approach is simple, yet effective and can be applied to any deep learning-based architecture which uses convolutional and batch normalization layers. Code is available at: https://github.com/aleemsidra/ConvLoRA.

ConvLoRA and AdaBN based Domain Adaptation via Self-Training

TL;DR

This work tackles unsupervised multi-target domain adaptation for medical image segmentation under cross-site shifts. It introduces ConvLoRA, a CNN-specific low-rank adapter that freezes pre-trained weights and learns a compact update with rank , together with Adaptive Batch Normalization (AdaBN) that uses target-domain statistics. The method is integrated into a UDAS-based pipeline with self-training, training only the ConvLoRA parameters and achieving substantial parameter reductions (≈99% fewer trainable parameters) while delivering competitive or superior segmentation performance across multiple target domains. The results demonstrate strong practical impact for MTDA in medical imaging, enabling efficient adaptation across diverse centers without maintaining separate full-model copies per target domain. Key ideas are applicable to CNN architectures with BN layers and are supported by extensive ablations on adapter placement, BN use, and training duration.

Abstract

Existing domain adaptation (DA) methods often involve pre-training on the source domain and fine-tuning on the target domain. For multi-target domain adaptation, having a dedicated/separate fine-tuned network for each target domain, that retain all the pre-trained model parameters, is prohibitively expensive. To address this limitation, we propose Convolutional Low-Rank Adaptation (ConvLoRA). ConvLoRA freezes pre-trained model weights, adds trainable low-rank decomposition matrices to convolutional layers, and backpropagates the gradient through these matrices thus greatly reducing the number of trainable parameters. To further boost adaptation, we utilize Adaptive Batch Normalization (AdaBN) which computes target-specific running statistics and use it along with ConvLoRA. Our method has fewer trainable parameters and performs better or on-par with large independent fine-tuned networks (with less than 0.9% trainable parameters of the total base model) when tested on the segmentation of Calgary-Campinas dataset containing brain MRI images. Our approach is simple, yet effective and can be applied to any deep learning-based architecture which uses convolutional and batch normalization layers. Code is available at: https://github.com/aleemsidra/ConvLoRA.
Paper Structure (11 sections, 1 equation, 2 figures, 2 tables)

This paper contains 11 sections, 1 equation, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. ConvLoRA
  5. Adaptive Batch Normalization (AdaBN)
  6. ConvLoRA and AdaBN based UDA
  7. Experimental Setup
  8. Results and Analysis
  9. Ablations
  10. Conclusion
  11. Acknowledgment

Figures (2)

  • Figure 1: 2D U-Net with Early Segmentation Head (ESH) is pre-trained on the source domain. ConvLoRA adapters facilitate adaptation in the encoder, along with AdaBN throughout the network.
  • Figure 2: Qualitative Results for target domains of CC359 souza2018open. Columns from left to right correspond to input images, ground truth, source U-Net model, UDAS sheikh2022unsupervised, and our ConvLoRA + AdaBN. It can be seen that our proposed adaptation has the most visual similarity to the ground truth.