Enhancing Radiation Oncology Imaging with a Novel Variational Model Decomposition, Radon Transformation, and Kohonen Self-Organizing Map Denoising Framework 📝

Author: Hassan Bagher-Ebadian, Justine M. Cunningham, Anthony J. Doemer, Mohammad M. Ghassemi, Joshua P. Kim, Benjamin Movsas, Kundan S Thind 👨‍🔬

Affiliation: Michigan State University, Henry Ford Health 🌍

Abstract:

Purpose: Reduction of noise in medical images critically enables improved accuracy in delineating tumors and organs at risk, leading to more precise treatment planning and safer image-guided radiation therapy. This study proposes a novel, modality agnostic denoising framework that integrates Variational-Model-Decomposition (VMD), Radon-Transformation (RDT), and Kohonen-Self-Organizing-Map (KSOM) to suppress and regulate high-frequency noise components of images while preserving structural details to enhance diagnostic reliability and accuracy.
Methods: T1-3D-VIBE-DIXON-IN Magnetic Resonance (MR) Images (n=4) and Cone Beam Computed Tomography (CBCT) images (n=7) acquired using a Siemens-Magnetom-Free.Max 0.55T and HyperSight-Ethos, respectively, were selected for this study. This framework used RDT with the optimal number of projections (hr∼𝜋𝐷/Δ𝑥, where 𝐷 is the reconstruction diameter with Δx-resolution) to ensure adequate angular sampling, minimize aliasing, and satisfy the Nyquist-criterion. The resampled projections were decomposed into a series of intrinsic-modes (IM=25) using VMD. The KSOM (Topology-size:5X5=25) was then applied to VMD results to regularize the decomposition by selecting the optimal number of modes, effectively suppressing noise by regularizing high-frequency components. The denoised images were reconstructed by applying the inverse-RDT to the projections reconstructed from the retained modes and signal-to-noise ratios (SNRs) of the images were compared.
Results: The framework reduced noise while preserving anatomical details. Visual and quantitative evaluation confirmed enhanced clarity, and SNRs improved significantly (31.75±1.40dB to 36.80±1.05dB for MR; 26.23±4.28dB to 31.18±3.14dB for CBCT). KSOM selected optimal modes (IM 19, 21) at central-frequency-bandwidths of 0.041, 0.043 cycles/mm (~7%, ~9% of Nyquist frequencies), balancing noise reduction and detail preservation. This was confirmed by targeted high-frequency removal affecting primarily Rician/Gaussian noise, preserving anatomical details.
Conclusion: This VMD-RDT-KSOM framework significantly enhanced SNR for low-field MRI and CBCT. It is modality-agnostic, offering potential for broader applications in radiation oncology and beyond. Future research on larger datasets and additional modalities will validate its versatility.