Tumor likelihood estimation on MRI prostate data by utilizing k-Space information

Motivation: In MRI, k-Space data is currently not directly used in the classification process, even though it may hold valuable additional information. Goal(s): This study aims to assess whether MRI k-Space data can directly - without image reconstruction - and accurately predict prostate cancer likelihood, reducing the need for image-domain reconstructions and enabling higher undersampling rates. Approach: A principal component analysis based coil compression pipeline was developed to process MRI k-Space data, tested with varying undersampling factors to simulate accelerated scans. Results: The model achieved an AUROC of 86.1% with fully sampled data and 78.0% at 16x undersampling, demonstrating robust predictions without canonical reconstruction methods. Impact: This study enables faster, reliable MRI-based prostate cancer predictions by utilizing k-Space raw data. It opens new possibilities for real-time diagnostics and broader applications of raw MRI data across clinical imaging.