Problem formulation beats architecture by 290x in MRI slice interpolation

A new study from Shamit Savant, published on arXiv (2605.16476), investigates deep learning for interpolating missing MRI slices to double through-plane resolution — a critical issue since clinical MRI typically has much coarser through-plane than in-plane resolution, limiting diagnostic utility. The work evaluated five architectures: CNN, U-Net, two GAN variants, and a Denoising Diffusion Probabilistic Model (DDPM) on prostate imaging data.

The most striking finding is that problem formulation had dramatically more impact than architectural complexity. By simply reformulating the interpolation to use adjacent slices (i-1, i+1) rather than distant slices (i-2, i+2), the author achieved a 58% improvement in SSIM (structural similarity index) across all deterministic architectures. The U-Net emerged as the best performer with PSNR of 30.08 dB and SSIM of 0.898 — a 10.1% improvement over the linear interpolation baseline. Notably, the DDPM showed poor reconstruction quality due to a fundamental mismatch between its stochastic generation process and the deterministic reconstruction requirements of slice interpolation. The paper concludes that problem formulation can have up to 290x more impact than architectural sophistication in medical imaging tasks, a finding with profound implications for researchers and practitioners.