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

TL;DR

This work addresses the slow reconstruction bottleneck in MRI-based prostate cancer prediction by leveraging complex-valued k-Space data rather than relying solely on reconstructed image-domain magnitudes. The authors propose a preprocessing pipeline that sums averages, applies PCA-based coil compression on k-Space, and feeds magnitude/phase (image domain) along with real/imag (k-Space) channels into a ConvNeXt network; they compare this against GRAPPA-based reconstruction. Key findings show an AUROC of $86.1\% \pm 1.8\%$ for fully reconstructed data with k-Space at $x2$ undersampling, and a meaningful AUROC of $71.4\% \pm 2.9\%$ at $x16$ undersampling using PCA k-Space processing, highlighting the potential for near real-time diagnostics with substantial time savings. The results suggest that preserving raw MRI information can improve prediction stability at higher undersampling and reduce processing time, enabling faster clinical decision-making; future work includes fully complex-valued networks and application to additional sequences.

Abstract

We present a novel preprocessing and prediction pipeline for the classification of magnetic resonance imaging (MRI) that takes advantage of the information rich complex valued k-Space. Using a publicly available MRI raw dataset with 312 subject and a total of 9508 slices, we show the advantage of utilizing the k-Space for better prostate cancer likelihood estimation in comparison to just using the magnitudinal information in the image domain, with an AUROC of $86.1\%\pm1.8\%$. Additionally, by using high undersampling rates and a simple principal component analysis (PCA) for coil compression, we reduce the time needed for reconstruction by avoiding the time intensive GRAPPA reconstruction algorithm. By using digital undersampling for our experiments, we show that scanning and reconstruction time could be reduced. Even with an undersampling factor of 16, our approach achieves meaningful results, with an AUROC of $71.4\%\pm2.9\%$, using the PCA coil combination and taking into account the k-Space information. With this study, we were able to show the feasibility of preserving phase and k-Space information, with consistent results. Besides preserving valuable information for further diagnostics, this approach can work without the time intensive ADC and reconstruction calculations, greatly reducing the post processing, as well as potential scanning time, increasing patient comfort and allowing a close to real-time prediction.

Figures (6)

• Figure 1: Methodical overview of the classification pipeline via PCA. The complex valued MRI raw data 2D slices are summed over the averages and then fed into the PCA. The first PCA component is then split into its image domain magnitude and phase, as well as its k-Space real and imaginary part. With this stacked input, the ConvNeXts then predict if the slice should be classified with a Pi-RADS score larger than two.
• Figure 2: Pi-RADS label distribution of the FastMRI Prostate dataset, as well as the distribution after combining the classes.
• Figure 3: Comparion of the standard preprocessing pipeline and the proposed preprocessing pipeline. The standard pipeline includes regridding, GRAPPA reconstruction, followed by coil combination in the image domain and the adc map calculation. The phase information are lost. The proposed pipeline first performs regridding, followed by a summation and coil compression via PCA. This approach is less computational intensive and preserves the complex valued k-Space data.
• Figure 4: Example of Undersampling in image domain from a fully sampled sample of the FastMRI+ Dataset. Undersampling refers to the missing out of lines in the k-Space. While reducing the imaging time, this leads to artifacts in the image domain, as seen in the top rows. The image wraps in itself, depending on the undersampling factor.
• Figure 5: Comparison of the results with and without additional k-Space information on the "gold-standard" data, consisting of GRAPPA reconstructed and coil combined adc, b50 and b1000 maps in the image domain (Mag) with different undersampling factors.