TL;DR
This study compares a new deep learning-based CT reconstruction (DLR) algorithm against traditional methods (MBIR, HIR, FBP) for lower extremity CT angiography (CTA). DLR significantly improves image quality, reducing noise and enhancing contrast, leading to superior diagnostic accuracy, making it a promising tool for evaluating peripheral artery disease.
What Problem Does This Paper Address?
Lower extremity CT angiography (CTA) is a crucial non-invasive imaging technique for diagnosing peripheral artery disease (PAD). However, achieving high image quality in CTA, especially for visualizing small vessels and subtle lesions in the lower extremities, remains a challenge. Traditional CT reconstruction methods like filtered back projection (FBP), hybrid iterative reconstruction (HIR), and model-based iterative reconstruction (MBIR) have improved image quality over time, but still have limitations in terms of noise, spatial resolution, and lesion detectability. This paper addresses the need for improved image quality in lower extremity CTA by evaluating a novel deep learning-based reconstruction (DLR) algorithm, aiming to determine if it offers superior performance compared to existing methods.
Methods
The researchers conducted a retrospective study involving 50 patients who underwent lower extremity CTA. Images were reconstructed using four different algorithms:
- Deep Learning Reconstruction (DLR): Specifically, the "AiCE Body Sharp" algorithm from Canon Medical Systems. This algorithm is trained on high-quality MBIR datasets to differentiate noise from signals using deep convolutional neural networks.
- Model-Based Iterative Reconstruction (MBIR): "FIRST Body Sharp" from Canon Medical Systems.
- Hybrid-Iterative Reconstruction (HIR): "AIDR 3D" with kernel FC08 from Canon Medical Systems.
- Filtered Back Projection (FBP): With kernel FC08 from Canon Medical Systems.
They performed both quantitative and qualitative analyses.
- Quantitative analysis involved measuring image noise (standard deviation - SD), contrast-to-noise ratio (CNR), signal-to-noise ratio (SNR), and a "blur effect" metric to assess sharpness. Noise power spectrum (NPS) curves were also analyzed to understand noise characteristics.
- Qualitative analysis involved two radiologists independently evaluating subjective image quality based on noise, sharpness, and natural appearance using a 5-point Likert scale.
- Diagnostic accuracy was assessed by comparing CTA findings with digital subtraction angiography (DSA) as the gold standard.
Datasets
The study utilized retrospective CTA data from 50 patients collected between January and May 2021. No new datasets were introduced for training or benchmarking. The primary dataset was the patient cohort undergoing lower extremity CTA.
Key Insights and Ideas
- DLR excels in noise reduction: The most striking finding is DLR's superior ability to reduce image noise, as evidenced by significantly lower SD values and higher CNR and SNR across various anatomical structures (aorta, femoral artery, popliteal artery, liver, psoas muscle). This directly translates to clearer visualization of vessels and surrounding tissues.
- DLR offers a good balance of sharpness and noise reduction: While FBP might show good sharpness, it comes with high noise. DLR, on the other hand, achieves a favorable balance, preserving sharpness without the excessive noise seen in FBP. The "blur effect" metric suggests DLR's sharpness is comparable to MBIR and better than HIR, though FBP showed the best sharpness (likely due to its inherent higher noise).
- Subjective quality aligns with objective findings: Radiologists consistently rated DLR images as having the best quality in terms of low noise, subjective sharpness, and natural appearance, reinforcing the objective measurements.
- Improved diagnostic accuracy: DLR demonstrated the highest sensitivity (98.4%) and specificity (97.2%) for detecting significant stenosis, surpassing the other reconstruction algorithms. This suggests DLR can lead to more reliable diagnoses of PAD.
Key Results
- Objective Image Quality: DLR showed significantly higher CNR and SNR and significantly lower SD compared to MBIR, HIR, and FBP across most anatomical regions.
- Noise Characteristics: DLR had the lowest noise magnitude. NPS analysis indicated that DLR and MBIR images had more high-frequency noise components than HIR.
- Sharpness (Blur Effect): DLR's blur effect was significantly lower than HIR. For larger arteries, the trend of blur effect from low to high was FBP, MBIR, DLR, and HIR.
- Subjective Image Quality: DLR received the highest scores for image noise, subjective sharpness, and natural appearance from radiologists.
- Diagnostic Accuracy: DLR achieved the highest sensitivity (98.4%) and specificity (97.2%) for detecting stenosis, followed closely by MBIR and HIR.
Limitations and Future Work
The authors acknowledge several limitations:
- Routine Radiation Dose: The study only compared DLR at routine radiation doses. Future work should investigate DLR's performance at lower radiation doses, as previous studies suggest it can maintain image quality while reducing dose.
- Kernel Selection: The "body sharp" kernel used for DLR and MBIR might have influenced noise levels and CNR/SNR. Further studies should explore the impact of different kernel selections.
- Sparse-View CT: The findings might not directly apply to newer sparse-view CT imaging techniques combined with deep learning, requiring further investigation.
- Limited DSA Cohort: The study had a small number of patients who underwent DSA, limiting the ability to perform sub-grouped analyses of diagnostic performance based on lesion location.
Future research should focus on evaluating DLR at reduced radiation doses, exploring various kernel settings, and investigating its performance with advanced CT imaging techniques.
Summary updated