Channel-guided image-adaptive CNN for task-aware denoising in medical imaging

Abstract

Medical image denoising with convolutional neural networks (CNNs) is commonly formulated as a pixel-wise regression problem, without explicitly accounting for information needed by downstream clinical tasks. As a result, conventional CNN denoisers may suppress weak but task-relevant image features while reducing noise. We present a task-aware image-adaptive denoising framework that fine-tunes a pretrained Denoising Convolutional Neural Network (DnCNN) separately for each input image at inference time. Partial least squares channels learned for binary signal detection tasks are used to extract task-relevant features, and the network is adapted by using the channel responses. The proposed approach is evaluated for a lesion-detection task in breast computed tomography (CT) imaging, where the lesion is modeled as a Gaussian signal embedded in breast CT images with synthetically added noise. Compared with the fixed baseline DnCNN, the image-adaptive model improved signal detection performance of a numerical observer, demonstrating that channel-guided test-time adaptation can preserve signal-detection information more effectively than conventional pixel-wise denoising.