RPS 613 - Artificial intelligence (AI) revising the physics in medical imaging

Purpose:

With filtered back-projected image reconstruction, spatial resolution performance is not dependent on image contrast or noise/dose. Existing iterative methods have been shown to produce a spatial resolution that is dependent on image contrast or noise/dose. Here, we characterise the contrast and dose dependence of TrueFidelity (GE Healthcare), a CE marked new deep learning image reconstruction (DLIR) approach.

Methods and materials:

We imaged acrylic, bone, polyethylene, and air phantom inserts. We imaged at dose levels of 16, 8, and 4 mGy. Images were reconstructed using 6 methods: filtered back projection (FBP), 2 levels of a statistical iterative reconstruction (ASiR-V), and 3 levels of the vendor’s new deep learning (DLIR) approach. The ASiR-V level was chosen based on a vendor recommendation (AR50, 50%). The tasked-based modulation transfer function (MTF task) methodology was used to obtain contrast dependent spatial resolution for each insert.

Results:

The 50% and 10% MTF task values for all DLIR strengths were all comparable to FBP and all ASiR-V levels. The 10% MTF task at 8 mGy for FBP was 0.69/0.65/0.65/0.66 for poly/air/acrylic/bone, respectively. The 10% MTF task at 8 mGy for 50% ASiR-V was 0.7/0.65/0.68/0.66 for poly/air/acrylic/bone, respectively. The 10% MTF task at 8 mGy for medium-strength DLIR was 0.71/0.69/0.69/0.69 for poly/air/acrylic/bone, respectively. All reconstruction methods showed an expected decreased performance when the focal spot switched from medium to large between the 8/4 and 16 mGy levels, respectively.

Conclusion:

Unlike other advanced iterative CT algorithms, this deep learning method did not exhibit contrast or noise/dose dependencies with respect to spatial resolution.

Limitations:

Other abstracts address noise texture differences.

Ethics committee approval

n/a

Funding:

GE Healthcare.

Purpose:

For the past few years, “low dose” was the benchmark for image quality in CT and the introduction of iterative reconstruction (IR) allowed dose reductions by about 50% compared to filtered back projection (FBP) at the expense of image texture. In this study, we aimed to assess whether a novel deep learning (DL) reconstruction can preserve the FPB-like image texture at dose levels attained by IR.

Methods and materials:

Noise (SD), NPS, and MTF were measured on Catphan phantom derived images. The acquisition occurred on a 512-slice GE Revolution at 100kVp, 350 ms, 0.98 pitch, 40 mm collimation, and 1.25 mm thickness. Two different dose levels (3.0 and 7.6mGy CTDIvol) were used, relevant for clinical practice and reconstructed with FBP, ASIR-V (50% and 100% blending), and DL.

Results:

Compared to FBP, both ASIR-V and DL reduced image noise. At 3.0mGy, SD decreased by 37%, 69%, and 35% and NPS area by 53%, 83%, and 54% using 50% ASIR-V, 100% ASIR-V, and DL, respectively. While the NPS apex of 50% and 100% ASIR-V (fpeak=0.23, 0.13 mm-1; favg=0.28, 0.18 mm-1) shifted towards lower frequencies with respect to FBP (fpeak=0.28 mm-1; favg=0.33 mm-1), no difference was observed with DL (fpeak=0.28 mm-1; favg=0.31 mm-1).

Noise and NPS of 3.0mGy DL (SD=13.5; fpeak=0.28 mm-1; favg=0.31 mm-1) and 7.6mGy FBP (SD=13.9; fpeak=0.30 mm-1; favg=0.3 3mm-1) were similar, suggesting a 60% dose reduction is possible with DL, while preserving the image texture.

No spatial frequency differences were observed between FBP, ASIR-V, and DL (MTF50%=0.36 mm-1).

Conclusion:

The DL-based reconstruction reduces noise with respect to FBP without modifying the image texture. Moreover, it has the potential to reduce the dose with respect to FBP, in a similar way as IR.

Limitations:

No clinical images were used to evaluate the radiologists’ perception of image quality and texture.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

New CT image reconstruction frameworks can alter the fundamental tradeoffs between spatial resolution, noise, and dose. Here, we investigate the performance of a new deep learning-based CT image reconstruction algorithm (TrueFidelity, GE Healthcare).

Methods and materials:

We imaged a phantom with low and high contrast inserts at a slice thickness ranging from 0.625 mm-5 mm. We imaged at dose levels of 16, 8, and 4 mGy. All measurements were repeated 5 times. Images were reconstructed using filtered back projection (FBP), 2 levels of an advanced statistical iterative reconstruction (ASiR-V), and 3 levels of a deep learning image reconstruction (DLIR) approach. The ASiR-V levels were chosen based on institutional (20%) and vendor (50%) recommended levels. CNR, uniformity, and CT number were assessed for each slice thickness, reconstruction type, and dose level.

Results:

All slice thicknesses, reconstruction types, and dose levels exhibited acceptable results for uniformity and CT number for all algorithms. The performance of CNR was highest for DLIR, followed closely by 50% ASiR-V, and was lowest for FBP. At 5 mm slice thickness and 16 mGy, DLIR low/medium/high CNR was 1.18/1.33/1.53 compared to 1.03 for FBP and 1.17/1.47 for 20%/50% ASiR-V. DLIR allowed for equal CNR as FBP at smaller slice thicknesses. For example, at 16 mGy, DLIR had the same CNR at 1.25 mm thickness as FBP had at 5 mm thickness.

Conclusion:

The new TrueFidelity DLIR method outperforms existing ASiR-V and FBP. Our results have motivated several sections in our institution to utilise thin 1.25 mm images for tasks for which they used to use a thicker 5 mm slice.

Limitations:

Additional abstracts address spatial resolution and noise texture.

Ethics committee approval

n/a

Funding:

GE Healthcare.

Purpose:

Deep learning image reconstruction (DLIR) is a new reconstruction method which can provide efficient noise reduction. The purpose of this phantom study was to investigate the relationship between radiation dose and physical metrics of the latest DLIR algorithm.

Methods and materials:

Acquisitions on a physical evaluation phantom (Catphan 700, The Phantom Laboratory, Salem, NY, USA) equipped with low- (⊿60H.U.) and high(⊿240H.U.)-contrast model objects were performed at 120kV and 10 dose levels (CTDIvol: 1-10mGy in 1mGy steps) with a multidetector CT scanner (Revolution CT, GE Healthcare, Milwaukee, WI, USA). Raw data was reconstructed with a 1.25 mm slice thickness using standard kernel for FBP, hybrid iterative reconstruction (HIR), and 3 strength levels of DLIR (low, med, high, TrueFidelity, GE Healthcare, Milwaukee, WI, USA). The noise power spectrum (NPS), modulation transfer function (MTF), and detectability index (SNR2) were computed.

Results:

The 50% MTF and 10% MTF of DLIR were comparable to those of FBP for low-contrast objects and were higher for high-contrast objects at all dose levels. The NPS peak frequency of DLIR was similar to FBF and was higher than HIR, indicating less noise texture change from FBP compared to HIR. The highest SNR2 was obtained with DLIR-high for both low- and high-contrast models. FBP-equivalent image quality can be achieved with -40% radiation dose for HIR, -50% for DLIR-low, -60% for DLIR-med, and -70% for DLIR-high for the high-contrast model. These values were -29%, -46%, -55%, and -64%, respectively, for the low-contrast model.

Conclusion:

DLIR can substantially reduce noise without sacrificing spatial resolution or changing noise texture, which might provide a dose reduction of up to -70%.

Limitations:

A phantom study

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To determine how radiation dose affects spatial resolution and noise in ultra-high-resolution CT (UHRCT) using deep learning reconstruction (DLR) and hybrid-iterative reconstruction (Hybrid-IR) algorithms.

Methods and materials:

We acquired images of a Catphan 500 phantom and a 320 mm water phantom using a UHRCT system with 0.25 mm effective detector size (Aquilion Precision, Canon Medical Systems) in high-resolution mode at CTDIvol values of 4.3, 9.1, and 22.3 mGy. The images were reconstructed with DLR (AiCE body standard) and Hybrid-IR (AiDR3D-enhanced, FC08). Modulation transfer functions (MTFs) at the isocentre were determined using the edge of the teflon rod of the catphan images. Image noise (standard deviation, SD) and noise power spectra (NPSs) were determined from the water phantom images, with the frequency where the NPS peaks used to quantify noise texture.

Results:

All results are provided in increasing dose order. The MTF of DLR is higher than that of Hybrid-IR at all dose-levels. The 50% of the MTFs is crossed at 9, 10, 10 lp/cm for DLR; Hybrid-IR: 6, 7, 7 lp/cm. SD values for DLR are 13.6, 18.7, and 17.1, while for Hybrid-IR these values are 21.9, 27.0, and 27.5. For both techniques, the NPS peak shifts with dose. Peak frequencies for DLR are 0.4, 0.7, and 0.9 lp/cm, while for Hybrid-IR these values are 0.7, 1.1, and 1.3 lp/cm.

Conclusion:

DLR creates higher-resolution images with lower noise compared to Hybrid-IR at all dose levels. As opposed to with filtered back projection, spatial resolution and noise coarseness are a function of the dose with these reconstruction algorithms.

Limitations:

The use of linear metrics with non-linear reconstruction algorithms must be performed with care and only for the evaluation of relative differences between techniques, as done here.

Ethics committee approval

n/a

Funding:

Funding received from Canon.

Purpose:

This work characterises the use of low kV imaging on Revolution Apex (GE Healthcare) with a new x-ray tube and TrueFidelity, a CE marked deep learning-based image reconstruction (DLIR) technique.

Methods and materials:

We imaged a CT phantom (30 by 40 cm solid water with 2, 5, and 15 mg/cc of iodine, fat, brain, and blood inserts) at 70, 80, 100, 120, and 140 kVp. We imaged at dose levels of 18, 8, and 4 mGy. Images were reconstructed using 3 methods: filtered back projection (FBP), a statistical iterative reconstruction (ASiR-V), and the vendor’s new deep learning image reconstruction approach. The ASiR-V levels were chosen based on institutional (20%) and vendor (50%) recommended levels. DLIR was set at medium strength. The relative dose reduction factor (RDF) methodology was applied to quantify the potential of changing kVp and or applying the ASiR-V/DLIR methods to reduce the dose.

Results:

As expected, for materials that increase the CT number while decreasing kVp, the RDF predicted a dose reduction for all recon types. For materials like the brain where FBP exhibited a dose penalty with decreasing kVp, DLIR enabled lowering of the kVp without any dose penalty (RDF for brain tissue FBP at 70 kVp was 1.10 while DLIR had a brain RDF at 70 kVp of 0.26). For 5 mg/cc strength iodine at 16 mGy and 70 kVp, FBP/ASiR-V 20%/ASiR-V 50%/DLIR had RDF values of 0.49/0.36/0.20/0.10, respectively.

Conclusion:

New deep learning reconstruction (TrueFidelity, GE Healthcare) allowed for a better simultaneous realisation of iodine and non-iodine imaging tasks on the same exam.

Limitations:

Noise texture and spatial resolution are evaluated in other abstracts.

Ethics committee approval

n/a

Funding:

GE Healthcare.

Purpose:

To assess the impact on image quality and dose reduction opportunities of a deep learning image reconstruction (DLIR) algorithm compared to a model-based iterative reconstruction (MBIR) algorithm.

Methods and materials:

Acquisitions were performed on an image quality phantom at 7 dose levels (CTDIvol: 15/10/7.5/5/2.5/1/0.5mGy). Raw data was reconstructed using the filtered back-projection (FBP), 2 levels of MBIR, and 3 levels of DLIR algorithms. The mean attenuation, noise, contrast-to-noise ratio (CNR), noise-power-spectrum (NPS), and task-based transfer function (TTF) were computed. The detectability index (d’) was computed to model the detection of a large mass in the liver, a small calcification, and small lesion with low-contrast.

Results:

CNR values obtained with all DLIR levels were lower than AV100 and higher than AV50 (except with DLIR-L). NPS peaks were higher with AV50 than with all DLIR levels and only higher with AV100 than with DLIR-L. The average NPS spatial frequencies were higher with DLIR than with MBIR. For all DLIR levels, TTF50% obtained with DLIR were higher than MBIR. d' values increased with the level of DLIR and the percentage of MBIR. d' were higher with DLIR than with AV50, but lower with DLIR-L and DLIR-M than with AV100. d' values were higher with DLIR-H than with AV100 for the small lesion with low-contrast (10%±4%) and in the same range for other simulated lesions.

Conclusion:

DLIR algorithms reduced noise and improved spatial resolution and detectability without changing the noise texture. Images obtained with DLIR seem to be more adapted to a dose optimisation process than those with MBIR.

Limitations:

Raw-data was reconstructed using the single-kernel “standard” and conducted on a phantom that does not take into account the difference in morphology between patients.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To reconstruct interventional tools from only 4 x-ray CBCT projections for tomographic interventional guidance.

Methods and materials:

Single or biplane x-ray fluoroscopy is the standard imaging technique for interventional guidance. Due to its projective nature, it only provides a limited ability to resolve three-dimensional structures. While tomographic (3D+time) interventional guidance could overcome this drawback, it is currently infeasible due to the exceedingly high dose to both the patient and the surgeon. Prior work demonstrated that interventional tools can be reconstructed from 10 to 20 cone-beam CT (CBCT) projections by using prior knowledge of an appropriately sampled patient scan together with principles from a compressed sensing theory. To achieve a further dose reduction, which is necessary to make tomographic interventional guidance clinically feasible, we trained a deep neural network to segment interventional tools (stents and guidewires) from CT reconstructions of only 4 x-ray projections. We assumed to have a perfect prior scan, therefore subtracted the prior scan from any interventional scan results in a sparse volume consisting of only interventional tools and noise. The network was trained on simulated CBCT data of several different stent models and guidewires and tested on both simulated data and measured data of a Siemens Zeego robot-driven C-arm system.

Results:

For both the simulated and the measured data, we observed very good agreement with the ground truth, indicating that the network is able to generalise from simulated to measured data while having been trained exclusively on the former.

Conclusion:

Deep learning-based reconstruction of interventional tools has the ability to overcome the dose issue of tomographic interventional guidance.

Limitations:

Currently, the method assumes to have a perfect patient prior. We want to address this limitation in the future.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

Detection of pulmonary nodules is a challenging task in low-dose CT. Dose reduction using iterative reconstruction (IR) results in image texture changes, impacting pathological structures. A novel deep learning (DL) reconstruction technique can suppress noise without impacting image texture, using dose levels comparable to IR. Our aim was to assess whether DL is non-inferior to IR for the clinical task of nodule detection.

Methods and materials:

Up to 6 (range 3-6, mean 4.2) artificial lung nodules (diameter: 3, 5, and 8mm; density: -800, -630, and +100HU) were inserted at different locations in the Lungman anthropomorphic phantom (Kyoto Kagaku). 6 configurations were imaged (10 abnormal and 6 normal). Each configuration was scanned at 7.6mGy (clinical protocol) and 3 different levels of dose reduction: 60% (3mGy), 80% (1.6mGy), and 95% (0.38mGy). Images were reconstructed using 2 different algorithms (50% ASIR-V and DL with low, medium, and high strength). The 256 CT-scans (160 abnormal, 96 normal) were evaluated by 4 chest radiologists. Nodules were located and scored on a 5-point scale, blinded for dose and reconstruction algorithm. Data was analysed by the jackknife free-response receiver operating characteristic method.

Results:

We found no statistically significant difference in nodule detection among the image reconstruction algorithms (p=0.987, the average across readers AUC: 0.555, 0.561, 0.557, and 0.558 for 50% ASIR-V, DL-L, DL-M, and DL-H, respectively) for all dose reduction levels together. When stratifying per dose, no statistical difference (p=0.97) was found for the lowest dose (0.38mGy).

Conclusion:

Our study suggests that this DL algorithm is non-inferior to IR for the clinical task of nodule detection, even at very low dose levels.

Limitations:

Research assessing the change in image texture has been conducted in a separate study to uncover the full potential of DL.

Ethics committee approval:

n/a

Funding:

No funding was received for this work.