Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WaRA: Wavelet Low Rank Adaptation

Moein Heidari, Yijin Huang, Yasamin Medghalchi, Alireza Rafiee, Roger Tam, Ilker Hacihaliloglu

Abstract

Adapting large pretrained vision models to medical image classification is often limited by memory, computation, and task-specific specializations. Parameter-efficient fine-tuning (PEFT) methods like LoRA reduce this cost by learning low-rank updates, but operating directly in feature space can struggle to capture the localized, multi-scale features common in medical imaging. We propose WaRA, a wavelet-structured adaptation module that performs low-rank adaptation in a wavelet domain. WaRA reshapes patch tokens into a spatial grid, applies a fixed discrete wavelet transform, updates subband coefficients using a shared low-rank adapter, and reconstructs the additive update through an inverse wavelet transform. This design provides a compact trainable interface while biasing the update toward both coarse structure and fine detail. For extremely low-resource settings, we introduce Tiny-WaRA, which further reduces trainable parameters by learning only a small set of coefficients in a fixed basis derived from the pretrained weights through a truncated SVD. Experiments on medical image classification across four modalities and datasets demonstrate that WaRA consistently improves performance over strong PEFT baselines, while retaining a favorable efficiency profile. Our code is publicly available at~\href{https://github.com/moeinheidari7829/WaRA}{\textcolor{magenta}{GitHub}}.

WaRA: Wavelet Low Rank Adaptation

Abstract

Adapting large pretrained vision models to medical image classification is often limited by memory, computation, and task-specific specializations. Parameter-efficient fine-tuning (PEFT) methods like LoRA reduce this cost by learning low-rank updates, but operating directly in feature space can struggle to capture the localized, multi-scale features common in medical imaging. We propose WaRA, a wavelet-structured adaptation module that performs low-rank adaptation in a wavelet domain. WaRA reshapes patch tokens into a spatial grid, applies a fixed discrete wavelet transform, updates subband coefficients using a shared low-rank adapter, and reconstructs the additive update through an inverse wavelet transform. This design provides a compact trainable interface while biasing the update toward both coarse structure and fine detail. For extremely low-resource settings, we introduce Tiny-WaRA, which further reduces trainable parameters by learning only a small set of coefficients in a fixed basis derived from the pretrained weights through a truncated SVD. Experiments on medical image classification across four modalities and datasets demonstrate that WaRA consistently improves performance over strong PEFT baselines, while retaining a favorable efficiency profile. Our code is publicly available at~\href{https://github.com/moeinheidari7829/WaRA}{\textcolor{magenta}{GitHub}}.

Paper Structure

This paper contains 12 sections, 7 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Tiny-WaRA for extreme low resource adaptation
  4. Results
  5. Datasets, Training and Evaluation Setup
  6. Comparisons with State-of-the-art
  7. Analysis and Ablation Study
  8. Wavelet energy across subbands.
  9. Singular value spectrum of the update.
  10. Shared vs. band-specific adapters across wavelet subbands
  11. Conclusion
  12. Acknowledgments

Figures (2)

  • Figure 1: Overview of WaRA and Tiny-WaRA. WaRA performs low-rank adaptation in a wavelet domain by applying DWT on patch tokens, updating subband coefficients with shared adapter parameters, and reconstructing an additive update via IDWT. Tiny-WaRA replaces the standard adapter with a compact parameterization that learns only a small coefficient vector in a fixed basis derived from pretrained weights balazy2024loramorris2026learning.
  • Figure 2: Complementary analysis of WaRA. (a): Wavelet subband energy pattern on Fundus. (b): Singular value spectrum of the corresponding weight update.