CycleINR: Cycle Implicit Neural Representation for Arbitrary-Scale Volumetric Super-Resolution of Medical Data

TL;DR

CycleINR reframes 3D medical volumetric super-resolution as continuous function approximation using a coordinate-based implicit neural representation, enabling arbitrary z-axis upsampling with a single model. It introduces a cycle-consistent loss to curb inter-slice over-smoothing and a Local Attention Mechanism to improve latent-code sampling, along with a new SNLI metric for inter-slice noise inconsistency. The method demonstrates competitive image quality while better preserving fine details and improving segmentation fidelity on the MSD liver dataset, showing practical impact for clinical review and downstream analysis. The combination of INR, cycle-consistency, and local attention yields flexible, efficient volumetric SR suitable for diverse clinical scenarios and potentially extendable to other modalities and temporal data.

Abstract

In the realm of medical 3D data, such as CT and MRI images, prevalent anisotropic resolution is characterized by high intra-slice but diminished inter-slice resolution. The lowered resolution between adjacent slices poses challenges, hindering optimal viewing experiences and impeding the development of robust downstream analysis algorithms. Various volumetric super-resolution algorithms aim to surmount these challenges, enhancing inter-slice resolution and overall 3D medical imaging quality. However, existing approaches confront inherent challenges: 1) often tailored to specific upsampling factors, lacking flexibility for diverse clinical scenarios; 2) newly generated slices frequently suffer from over-smoothing, degrading fine details, and leading to inter-slice inconsistency. In response, this study presents CycleINR, a novel enhanced Implicit Neural Representation model for 3D medical data volumetric super-resolution. Leveraging the continuity of the learned implicit function, the CycleINR model can achieve results with arbitrary up-sampling rates, eliminating the need for separate training. Additionally, we enhance the grid sampling in CycleINR with a local attention mechanism and mitigate over-smoothing by integrating cycle-consistent loss. We introduce a new metric, Slice-wise Noise Level Inconsistency (SNLI), to quantitatively assess inter-slice noise level inconsistency. The effectiveness of our approach is demonstrated through image quality evaluations on an in-house dataset and a downstream task analysis on the Medical Segmentation Decathlon liver tumor dataset.

Figures (3)

• Figure 1: The core concept of CycleINR involves the initial use of signal $\textbf{X}$ to fit a continuous Implicit Neural Representation (INR) function (shown in dotted blue curves). Subsequently, new points (depicted as hollow orange dots) are sampled from this function to create a new INR function (illustrated by dotted orange curves). The signal $\hat{\textbf{X}}$ is then sampled from the new function at the same positions as $\textbf{X}$. The construction of a cycle-consistent loss is achieved by assessing the similarity between $\hat{\textbf{X}}$ and $\textbf{X}$.
• Figure 2: Schematic plot of our CycleINR framework for volumetric super-resolution. (a) Meta architecture of the proposed CycleINR. (b) Local Attention Mechanism (LAM) enhanced INR model. (c) Attention-enhanced Latent Code Grid Sampling (ALCGS) process.
• Figure 3: Visualization results of different super-resolution (x5) methods from axial, sagittal, and coronal views. The second and third rows display newly generated images at new positions, while the first and fourth rows show generated images at their original positions. In the bottom row, red arrows in TVSRN, ArSSR, and SAINT highlight horizontal lines in the coronal view, indicating slice-wise inconsistency.