Impact of deep Learning-enhanced contrast on diagnostic accuracy in stroke CT angiography

Abstract

Purpose To examine the impact of deep learning-augmented contrast enhancement on image quality and diagnostic accuracy of poorly contrasted CT angiography in patients with suspected stroke. Methods This retrospective single-centre study included 102 consecutive patients who underwent CT imaging for suspected stroke between 01/2021 and 12/2022, including whole brain volume perfusion CT (VPCT) and, specifically, a poorly contrasted CT angiography (defined as < 350HU in the proximal MCA). CT angiography imaging data was reconstructed using i.) an iterative reconstruction kernel (conventional CTA, c-CTA) as well as ii.) an iodine-based contrast boosting deep learning model (Deep Learning-enhanced CTA, DLe-CTA). For quantitative analysis, the slope, contrast-to-noise ratio (CNR), and signal-to-noise ratio (SNR) were determined. Qualitative image analysis was conducted by three readers, rating image quality and vessel-specific parameters on a 4-point Likert scale. Readers evaluated both datasets for cerebral vessel occlusion presence. VPCT served as the reference standard for calculating sensitivity and specificity. Results 102 patients were evaluated (mean age 69 ± 13 years; 70 men). DLe-CTA outperformed c-CTA in quantitative (all items p < 0.001) and qualitative image analysis (all items p < 0.05). VPCT revealed 58/102 patients with vascular occlusion. DLe-CTA resulted in significantly higher sensitivity compared to c-CTA (p < 0.001); (all readers put together: c-CTA: 142/174 [81.6 %; 95 % CI: 75.0 %-87.1 %] vs. DLe-CTA 163/174 [94 %; 95 % CI: 89.0 %-96.8 %]). One false positive finding occurred on DLe-CTA (specificity 1/132) [99.2 %; 95 % CI: 95.9 %-100 %]. Conclusions Deep learning-augmented contrast enhancement improves the image quality and increases the sensitivity of detection vessel occlusions in poorly contrasted CTA.