Impact of Deep Learning-Based Image Reconstruction on Tumor Visibility and Diagnostic Confidence in Computed Tomography.

Abstract

Deep learning image reconstruction (DLIR) has shown potential to enhance computed tomography (CT) image quality, but its impact on tumor visibility and adoption among radiologists with varying experience levels remains unclear. This study compared the performance of two deep learning-based image reconstruction methods, DLIR and Pixelshine, an adaptive statistical iterative reconstruction-volume (ASIR-V) method, and filtered back projection (FBP) across 33 contrast-enhanced CT staging examinations, evaluated by 20-24 radiologists. The signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were measured for tumor and surrounding organ tissues across DLIR (Low, Medium, High), Pixelshine (Soft, Ultrasoft), ASIR-V (30-100%), and FBP. In two blinded surveys, radiologists ranked eight reconstructions and assessed four using a 5-point Likert scale in arterial and portal venous phases. DLIR consistently outperformed other methods in SNR, CNR, image quality, image interpretation, structural differentiability and diagnostic certainty. Pixelshine performed comparably only to ASIR-V 50%. No significant differences were observed between junior and senior radiologists. In conclusion, DLIR-based techniques have the potential to establish a new benchmark in clinical CT imaging, offering superior image quality for tumor staging, enhanced diagnostic capabilities, and seamless integration into existing workflows without requiring an extensive learning curve.