Deep Learning Super-Resolution from Normal to Ultra-High Resolution CT: Conditional Diffusion Model Development and...

To develop a Conditional Denoising Diffusion Probabilistic Model (Conditional DDPM) for super-resolution (SR) of normal resolution (NR) multi-detector CT images (~0.4 mm detail size) to a level consistent with the recently introduced ultra-high-resolution (UHR) CT (~0.2 mm detail size) and to evaluate the impact of SR on texture metrics of trabecular bone.

Four human cadaver femurs were imaged using Canon Precision CT in an NR mode, with spatial resolution representative of the current conventional multi-detector CT (0.25 × 0.25 × 0.5 mm voxels), and in a novel UHR mode (0.125 × 0.125 × 0.25 mm voxels). For training the conditional DDPM, 7717 spatially-aligned patches (16 × 16 mm) were extracted from the NR and UHR images of two femurs; the NR patch was used as a condition, the UHR patch was the target. Model validation involved 5400 patches from the third femur. Testing was performed by slice-by-slice application of the trained SR on 190 cubic regions of interest (ROIs) from both ends of the fourth femur. The resulting SR images of trabecular bone were evaluated qualitatively and in terms of agreement of radiomic texture metrics with UHR data. Specifically, 24 Grey Level Co-occurrence Matrix (GLCM) features were extracted within 5 mm sphere masks in corresponding cancellous bone regions of the UHR, SR, and NR datasets. Concordance correlation coefficients (CCC) were calculated for NR vs. UHR and SR vs. UHR to assess UHR texture restoration using SR. Additionally, we investigated the overlap of NR vs. UHR and SR vs. UHR texture in a 2-dimensional space span by the first two principal components of ROI GLCM features.

Visually, SR using the conditional DDPM appeared to enhance the resolution of NR images (condition) to a level comparable to UHR CT. GLCM matrix visualization confirmed that the SR images resembled the GLCM matrix of UHR CT more closely than NR. Notably, the SR ROIs achieved significantly higher CCCs against UHR for 22 texture features, with an average improvement of 161.3% and a maximum increase of 465.5% for the Inverse Difference Moment. However, two metrics, Cluster Prominence and Cluster Shade, exhibited slightly lower CCCs for SR vs. UHR compared to NR vs. UHR, with reductions of 6.4% and 4.3%, respectively. PCA visualization further confirmed a greater trabecular bone texture feature overlap between SR and UHR compared to NR and UHR.

The conditional DDPM successfully enhances NR CT images to UHR quality, enabling more accurate visualization and quantification of bone microarchitecture. Importantly for bone radiomic applications, SR trabecular image texture features agree well with UHR CT for the majority of examined features, which may enable application of SR to harmonize NR data with UHR for predictive model development.