Task-Aware Enhancement of Low-SNR Prostate DWI on the FastMRI Prostate Dataset
Abstract
Prostate diffusion-weighted imaging (DWI) is highly vulnerable to low signal-to-noise ratio (SNR), which can degrade image quality and obscure diagnostically important tissue characteristics. This paper proposes a task-aware enhancement framework for low-SNR prostate DWI using the public fastMRI Prostate dataset. The method operates on reconstructed magnitude images derived from multi-coil raw MRI data and is trained with a hybrid objective that jointly enforces image restoration, edge preservation, and preservation of information relevant to prostate cancer prediction. Unlike conventional denoising methods that optimize visual fidelity alone, the proposed framework incorporates auxiliary task-aware supervision based on prostate cancer likelihood labels to preserve diagnostically useful content during enhancement. The method is evaluated under controlled low-SNR conditions using both image-quality metrics and downstream classification performance. On the fastMRI Prostate test set, the full model achieved the best diagnostic performance (AUC=0.761, F1=0.719) while remaining competitive in image restoration quality (PSNR=32.53 dB, SSIM=0.868), demonstrating a more favorable trade-off between visual fidelity and diagnostic preservation than appearance-only enhancement.