Authors

Wentao Chenand Weimin Zhou

Abstract

Image denoising is a fundamental inverse problem in medical imaging. Traditional sparsity-promoting approaches, such as penalized least-squares (PLS) with total variation (TV), have been widely employed to address such inverse problems. However, the handcrafted regularizers used in these methods cannot fully model complex image statistics, often leading to suboptimal image quality (IQ) performance. Supervised learning-based methods employing convolutional neural networks (CNNs) have emerged as powerful denoising approaches that can capture complex image statistics and improve IQ. However, CNNs trained with loss functions based solely on traditional IQ measures may remove subtle image details that are critical for task performance (e.g., tumor detection accuracy), resulting in degraded task-based IQ measures. In this work, we propose a novel task-based image-adaptive strategy to fine-tune the pre-trained CNN denoiser for each individual image by further promoting data fidelity while penalizing deviations in test statistics of a model observer for classification tasks. Numerical studies involving a stochastic binary texture model are conducted, and it is demonstrated that the proposed method achieves substantial improvements compared to previous medical image denoising methods in both traditional and task-based IQ metrics.

Article

pdf

BibTeX

@article{wentaochen2026,
  title={Medical Image Denoising with Task-based Image-adaptive Denoising CNN},
  author={Wentao Chen and Weimin Zhou},
  journal={Medical Imaging 2026: Image Perception, Observer Performance, and Technology Assessment},
  year={2026},
  month={}
}