AMeta-FD: Adversarial Meta-learning for Few-shot retinal OCT image Despeckling

Area of Science:

• Biomedical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Speckle noise in Optical Coherence Tomography (OCT) images degrades image quality, hindering crucial analysis tasks like retinal layer segmentation.
• Deep learning offers a robust and cost-effective alternative to traditional image processing and hardware solutions for noise reduction.
• Acquiring large annotated datasets for training deep learning models in medical imaging is often time-consuming and resource-intensive.

Purpose of the Study:

• To propose a novel deep learning method, Adversarial Meta-learning for Few-shot raw retinal OCT image Despeckling (AMeta-FD), for efficient speckle noise reduction in OCT images.
• To develop a method that requires significantly less training data compared to conventional deep learning approaches.
• To introduce a new suppression loss function to improve the effectiveness of noise reduction, particularly in non-tissue regions.

Main Methods:

• Adversarial meta-training on synthetic noisy OCT image pairs followed by fine-tuning on a small dataset of raw-clean OCT images.
• Introduction of a novel suppression loss function to minimize the impact of non-tissue pixels during the despeckling process.
• Generation of ground truth data by registering and averaging multiple repeated OCT image acquisitions.

Main Results:

• AMeta-FD achieves performance comparable to traditional transfer learning methods that utilize the entire training dataset, despite requiring only 12% of the data (60 raw-clean image pairs).
• The proposed method demonstrates a significant improvement in signal-to-noise ratio (SNR), surpassing traditional non-learning-based despeckling methods by at least 15 dB.
• AMeta-FD outperforms recent meta-learning-based denoising methods, including Few-Shot Meta-Denoising (FSMD) by 11.01 dB and previous best methods by 3 dB.

Conclusions:

• AMeta-FD presents a highly effective and data-efficient solution for speckle noise reduction in retinal OCT images.
• The method's ability to achieve state-of-the-art performance with limited data makes it a valuable tool for clinical applications where data acquisition is challenging.
• The novel suppression loss and adversarial meta-learning strategy contribute to the superior performance and robustness of AMeta-FD.