Evaluation of Coronary CT with Super-Resolution Precise IQ Engine (PIQE)
Fuminari Tatsugami, MD, PhD
This lecture presents a brief report on the current status of Deep Learning Reconstruction (DLR) followed by a discussion of the author's actual clinical experience with Canon’s Super-Resolution DLR known as Precise IQ Engine (PIQE), which is a next-generation DLR method.
Current status of DLR
The CT systems manufactured by Canon Medical Systems Corporation employ four basic image reconstruction techniques. These are the filtered back projection technique, which has been used in clinical practice for many years, as well as three other techniques: (1) Adaptive Iterative Dose Reduction 3D (AIDR 3D), which is a hybrid IR approach, (2) Forward projected model-based Iterative Reconstruction SoluTion (FIRST), which is a model-based IR approach, and (3) Advanced intelligent Clear-IQ Engine (AiCE), which is a DLR approach. Among these techniques, AIDR 3D has gained the most widespread clinical acceptance. Its main advantages are reduced image noise and shorter image reconstruction times. FIRST is able to provide high-resolution images, but the image reconstruction time is longer.
In the AiCE technique, which was developed in 2018, an algorithm is obtained by training a deep convolutional neural network (DCNN) using high-quality training images obtained using FIRST and low-quality images created by intentionally degrading the high-quality training images by adding artificial noise. When the algorithm obtained by Deep Learning is employed in clinical practice, high-quality images can be generated from images acquired by low-dose scanning. When AiCE in used in coronary CT imaging with a standard radiation dose, images with lower noise levels and a higher CNR can be obtained. Image quality is markedly improved, with much clearer depiction of the vessel walls. Nevertheless, the main objective of AiCE is noise reduction, so there is also a need to develop higher resolution image reconstruction methods. This is why Canon Medical Systems Corporation decided to begin development of Super- Resolution PIQE.
In the AiCE technique, which was developed in 2018, an algorithm is obtained by training a deep convolutional neural network (DCNN) using high-quality training images obtained using FIRST and low-quality images created by intentionally degrading the high-quality training images by adding artificial noise. When the algorithm obtained by Deep Learning is employed in clinical practice, high-quality images can be generated from images acquired by low-dose scanning. When AiCE in used in coronary CT imaging with a standard radiation dose, images with lower noise levels and a higher CNR can be obtained. Image quality is markedly improved, with much clearer depiction of the vessel walls. Nevertheless, the main objective of AiCE is noise reduction, so there is also a need to develop higher resolution image reconstruction methods. This is why Canon Medical Systems Corporation decided to begin development of Super- Resolution PIQE.
Next-generation DLR PIQE
1. Outline of Super-Resolution PIQE
Super Resolution refers to a special processing technology that restores the loss of resolution that occurs during image processing. This is not true for 512 matrix but also improves the visualisation fine structures.
To achieve the highest image resolution, it is still necessary to use the Aquilion Precision Ultra-High Resolution CT system, also manufactured by Canon Medical Systems. Aquilion Precision features a detector with 0.25 mm × 160 rows, 1792 channels, a minimum focal spot size of 0.4 mm × 0.5 mm, and a spatial resolution of 0.15 mm. These features make it possible to obtain much more precise structural information than with a conventional CT system. Furthermore, by employing FIRST or AiCE in SHR mode, high-resolution images that are impossible to acquire using a conventional CT system can be obtained, allowing fine structures to be visualised even more clearly.
To develop PIQE, training data is prepared by down-sampling Precision Ultra High-Resolution data. The DCNN is trained using actual UHR data paired with down-sampled, simulated Normal Resolution data. Once trained, the network is validated and applied to the Aquilion ONE scanner where it does not continue to learn. (Figure 1)
2. Image quality evaluation
The noise characteristics (NPS), image noise (SD), and spatial resolution (MTF) of AIDR 3D, FIRST, AiCE, and PIQE were evaluated and compared.
NPS was evaluated using a water phantom. PIQE showed the greatest noise magnitude reduction and was the most effective algorithm at reducing low-frequency, large-grain noise. (Figure 2)
Super Resolution refers to a special processing technology that restores the loss of resolution that occurs during image processing. This is not true for 512 matrix but also improves the visualisation fine structures.
To achieve the highest image resolution, it is still necessary to use the Aquilion Precision Ultra-High Resolution CT system, also manufactured by Canon Medical Systems. Aquilion Precision features a detector with 0.25 mm × 160 rows, 1792 channels, a minimum focal spot size of 0.4 mm × 0.5 mm, and a spatial resolution of 0.15 mm. These features make it possible to obtain much more precise structural information than with a conventional CT system. Furthermore, by employing FIRST or AiCE in SHR mode, high-resolution images that are impossible to acquire using a conventional CT system can be obtained, allowing fine structures to be visualised even more clearly.
To develop PIQE, training data is prepared by down-sampling Precision Ultra High-Resolution data. The DCNN is trained using actual UHR data paired with down-sampled, simulated Normal Resolution data. Once trained, the network is validated and applied to the Aquilion ONE scanner where it does not continue to learn. (Figure 1)
2. Image quality evaluation
The noise characteristics (NPS), image noise (SD), and spatial resolution (MTF) of AIDR 3D, FIRST, AiCE, and PIQE were evaluated and compared.
NPS was evaluated using a water phantom. PIQE showed the greatest noise magnitude reduction and was the most effective algorithm at reducing low-frequency, large-grain noise. (Figure 2)
Figure 1. Outline of PIQE
Next, SD was measured in the ascending aorta, left atrium, and interventricular septum. PIQE showed better noise reduction capabilities than AiCE (Figure 3).
MTF is normally measured using a phantom, but in the present evaluation, measurement was performed using the edge method for clinical images at the margin between the left ventricular lumen and the myocardium. The profile curve in a direction orthogonal to the edge was measured, and the MTF of each reconstruction image was then calculated. Figure 4 shows a comparison of the MTF for each reconstruction technique (left), in addition to graphs for 2% MTF (upper right) and 10% MTF (lower right). PIQE is seen to have higher spatial resolution than FIRST.
MTF is normally measured using a phantom, but in the present evaluation, measurement was performed using the edge method for clinical images at the margin between the left ventricular lumen and the myocardium. The profile curve in a direction orthogonal to the edge was measured, and the MTF of each reconstruction image was then calculated. Figure 4 shows a comparison of the MTF for each reconstruction technique (left), in addition to graphs for 2% MTF (upper right) and 10% MTF (lower right). PIQE is seen to have higher spatial resolution than FIRST.
Figure 2. Comparison of NPS
Figure 3. Comparison of SD (n=15)
Figure 4. Comparison of MTF
Presentation of coronary CT images
Figure 5 shows eccentric plaque in the proximal left anterior descending (LAD) artery. Compared to the other reconstruction techniques, PIQE (d) depicts the vessels more clearly, allowing the locations of the plaques to be identified with greater precision (red arrowheads).
Figure 6 shows coronary artery stents. The struts of the stents are more clearly visualised with PIQE (d) and it can be seen that the vascular lumen is wider. The profiles generated in the same segment demonstrate that PIQE provides higher resolution than the other techniques.
Figure 7 shows MPR images of the distal branches of the coronary arteries. The distal branches are more clearly depicted with PIQE (d) than with AIDR (a) or AiCE (c) (red circles). FIRST (b) provides clear visualisation, but is still not able to match PIQE.
Figure 8 shows an old myocardial infarction with a small area of infarction at the apex (yellow arrowheads). The infarction can be seen in all of the images, but noise levels are significantly lower with PIQE (d), allowing the area of infarction to be more clearly visualised.
Figure 9 shows images of an obese patient with a BMI of 38.7. Image quality is significantly improved with PIQE (d). The SD is as low as 20.8 HU, compared to 45.0 HU for AIDR 3D (a). Image quality is also improved in the curved MPR of the coronary arteries with PIQE, providing clear images suitable for diagnosis.
The time required for image reconstruction after the acquisition of 640 cardiac CT slices was approximately 20 seconds for AIDR 3D, approximately 45 seconds for AiCE, approximately one minute and 45 seconds for PIQE, and approximately four minutes and 30 seconds for FIRST. The time required for reconstruction with PIQE is only one-third of that achieved with FIRST.
Figure 6 shows coronary artery stents. The struts of the stents are more clearly visualised with PIQE (d) and it can be seen that the vascular lumen is wider. The profiles generated in the same segment demonstrate that PIQE provides higher resolution than the other techniques.
Figure 7 shows MPR images of the distal branches of the coronary arteries. The distal branches are more clearly depicted with PIQE (d) than with AIDR (a) or AiCE (c) (red circles). FIRST (b) provides clear visualisation, but is still not able to match PIQE.
Figure 8 shows an old myocardial infarction with a small area of infarction at the apex (yellow arrowheads). The infarction can be seen in all of the images, but noise levels are significantly lower with PIQE (d), allowing the area of infarction to be more clearly visualised.
Figure 9 shows images of an obese patient with a BMI of 38.7. Image quality is significantly improved with PIQE (d). The SD is as low as 20.8 HU, compared to 45.0 HU for AIDR 3D (a). Image quality is also improved in the curved MPR of the coronary arteries with PIQE, providing clear images suitable for diagnosis.
The time required for image reconstruction after the acquisition of 640 cardiac CT slices was approximately 20 seconds for AIDR 3D, approximately 45 seconds for AiCE, approximately one minute and 45 seconds for PIQE, and approximately four minutes and 30 seconds for FIRST. The time required for reconstruction with PIQE is only one-third of that achieved with FIRST.
Figure 5. Visualisation of plaque in the coronary arteries
Figure 6. Visualisation of coronary artery stents
Figure 7. Visualisation of distal branches of the coronary arteries
Figure 8. Visualisation of old myocardial infarction
Figure 9. Improved image quality in an obese patient (BMI: 38.7)
Summary
PIQE is a new image reconstruction technique for obtaining high-resolution images based on Deep Learning technology. In coronary CT, PIQE is an advanced image reconstruction technique that shares the advantages of AiCE, but with higher image resolution and lower noise levels than AiCE. //
* The contents of this report include the personal opinions of the author based on his clinical experience and knowledge.
Fuminari Tatsugami, MD, PhD
Department of Diagnostic Radiology, Hiroshima University Hospital, Japan
Department of Diagnostic Radiology, Hiroshima University Hospital, Japan
Due to medical device regulatory reasons, not all products/service displayed on this Canon Medical Systems ANZ webpage are available in all countries, regions or markets. Future availability of the products/service cannot also be guaranteed. Please contact your local Canon Medical Systems representative for further details.
