Towards Texture Accurate Slice Interpolation of Medical Images Using PixelMiner

Quantitative analysis models are used for various medical imaging tasks such as registration, classification, object detection, and segmentation which all benefit from high-quality imaging. We propose PixelMiner, a convolution-based deep-learning model which interpolates computed tomography (CT) imaging slices while better preserving quantitative imaging features than standard interpolation methods. PixelMiner was designed to produce texture-accurate slice interpolations by trading off pixel accuracy for texture accuracy using a novel architecture. PixelMiner was trained on a large dataset consisting of 7092 lung CT scans and validated using external lung CT datasets. We demonstrated the model's effectiveness by using edge preservation ratio (EPR), compared texture features commonly used in radiomics, and performed a qualitative assessment by human trial. PixelMiner had the highest EPR 82% (p<.01) of the time compared to the closest competing method. It had the lowest texture error, using a normalized root mean squared error (NRMSE) of 0.11 (p<.01), with the highest reproducibility of concordance correlation coefficient (CCC) ≥ 0.85 (p<.01). PixelMiner was chosen 72% of the time by human evaluation (p<.01). PixelMiner was not only demonstrated quantitatively to have improved structural and textural constructions but also shown to be preferable qualitatively.