NR-MRC: A No-Reference Magnetic Resonance Curve Estimation Network for Tumor MR Image Quality Assessment and...

Accurate image quality assessment (IQA) is essential for evaluating tumor magnetic resonance (MR) images in clinical practice and computer-aided diagnosis. However, most existing IQA models are designed for natural images and require reference images for training. Variability in tissue characteristics, lesion location, and lesion morphology, together with the scarcity of high-quality MR reference images, makes quality assessment of tumor MR images particularly challenging.

To address these challenges, we propose a no-reference magnetic resonance curve estimation network named NR-MRC. First, a magnetic resonance curve estimation module uses nonlinear curve mapping to estimate spatially adjusted intensity and detail features without requiring reference tumor MR images. Second, a multi-scale feature estimation module extracts contextual features at different scales and fuses them with the initial estimates to generate enhanced MR images. Finally, a magnetic resonance scene statistics blind quality model evaluates feature consistency between the original images and their enhanced counterparts.

Experiments on publicly available brain tumor and bladder cancer datasets demonstrate that NR-MRC outperforms the compared IQA and image enhancement methods. In the tested U-Net and DeepLab v3 settings, training with images selected by high NR-MRC scores or enhanced by NR-MRC increased segmentation metrics by approximately 20%. These task-specific gains do not by themselves establish improved clinical diagnostic image quality.

NR-MRC provides no-reference quality assessment and enhancement for tumor MR images and shows potential utility in the evaluated downstream segmentation tasks.