Artificial intelligence-based PET denoising could allow a two-fold reduction in [18F]FDG PET acquisition time in digital PET/CT

The impact of introducing deep learning based [18F]FDG PET denoising on EORTC and PERCIST therapeutic response assessments in digital PET/CT

BACKGROUND

[18F]FDG PET denoising by SubtlePET™ using deep learning artificial intelligence (AI) was previously found to induce slight modifications in lesion and reference organs' quantification and in lesion detection. As a next step, we aimed to evaluate its clinical impact on [18F]FDG PET solid tumour treatment response assessments, while comparing "standard PET" to "AI denoised half-duration PET" ("AI PET") during follow-up.

RESULTS

110 patients referred for baseline and follow-up standard digital [18F]FDG PET/CT were prospectively included. "Standard" EORTC and, if applicable, PERCIST response classifications by 2 readers between baseline standard PET1 and follow-up standard PET2 as a "gold standard" were compared to "mixed" classifications between standard PET1 and AI PET2 (group 1; n = 64), or between AI PET1 and standard PET2 (group 2; n = 46). Separate classifications were established using either standardized uptake values from ultra-high definition PET with or without AI denoising (simplified to "UHD") or EANM research limited v2 (EARL2)-compliant values (by Gaussian filtering in standard PET and using the same filter in AI PET). Overall, pooling both study groups, in 11/110 (10%) patients at least one EORTCUHD or EARL2 or PERCISTUHD or EARL2 mixed vs. standard classification was discordant, with 369/397 (93%) concordant classifications, unweighted Cohen's kappa = 0.86 (95% CI: 0.78-0.94). These modified mixed vs. standard classifications could have impacted management in 2% of patients.

CONCLUSIONS

Although comparing similar PET images is preferable for therapy response assessment, the comparison between a standard [18F]FDG PET and an AI denoised half-duration PET is feasible and seems clinically satisfactory.

Nuclear medicine technologists practice impacted by AI denoising applications in PET/CT images

PURPOSE

Artificial intelligence (AI) in positron emission tomography/computed tomography (PET/CT) can be used to improve image quality when it is useful to reduce the injected activity or the acquisition time. Particular attention must be paid to ensure that users adopt this technological innovation when outcomes can be improved by its use. The aim of this study was to identify the aspects that need to be analysed and discussed to implement an AI denoising PET/CT algorithm in clinical practice, based on the representations of Nuclear Medicine Technologists (NMT) from Western-Switzerland, highlighting the barriers and facilitators associated.

METHODS

Two focus groups were organised in June and September 2023, involving ten voluntary participants recruited from all types of medical imaging departments, forming a diverse sample of NMT. The interview guide followed the first stage of the revised model of Ottawa of Research Use. A content analysis was performed following the three-stage approach described by Wanlin. Ethics cleared the study.

RESULTS

Clinical practice, workload, knowledge and resources were de 4 themes identified as necessary to be thought before implementing an AI denoising PET/CT algorithm by ten NMT participants (aged 31-60), not familiar with this AI tool. The main barriers to implement this algorithm included workflow challenges, resistance from professionals and lack of education; while the main facilitators were explanations and the availability of support to ask questions such as a "local champion".

CONCLUSION

To implement a denoising algorithm in PET/CT, several aspects of clinical practice need to be thought to reduce the barriers to its implementation such as the procedures, the workload and the available resources. Participants emphasised also the importance of clear explanations, education, and support for successful implementation.

IMPLICATIONS FOR PRACTICE

To facilitate the implementation of AI tools in clinical practice, it is important to identify the barriers and propose strategies that can mitigate it.

Artificial intelligence solution to accelerate the acquisition of MRI images: Impact on the therapeutic care in oncology in radiology and radiotherapy departments

PURPOSE

MRI is essential in the management of brain tumours. However, long waiting times reduce patient accessibility. Reducing acquisition time could improve access but at the cost of spatial resolution and diagnostic quality. A commercially available artificial intelligence (AI) solution, SubtleMR™, can increase the resolution of acquired images. The objective of this prospective study was to evaluate the impact of this algorithm that halves the acquisition time on the detectability of brain lesions in radiology and radiotherapy.

MATERIAL AND METHODS

The T1/T2 MRI of 33 patients with brain metastases or meningiomas were analysed. Images acquired quickly have a matrix divided by two which halves the acquisition time. The visual quality and lesion detectability of the AI images were evaluated by radiologists and radiation oncologist as well as pixel intensity and lesions size.

RESULTS

The subjective quality of the image is lower for the AI images compared to the reference images. However, the analysis of lesion detectability shows a specificity of 1 and a sensitivity of 0.92 and 0.77 for radiology and radiotherapy respectively. Undetected lesions on the IA image are lesions with a diameter less than 4mm and statistically low average gadolinium-enhancement contrast.

CONCLUSION

It is possible to reduce MRI acquisition times by half using the commercial algorithm to restore the characteristics of the image and obtain good specificity and sensitivity for lesions with a diameter greater than 4mm.

Ultra-fast whole-body bone tomoscintigraphies achieved with a high-sensitivity 360° CZT camera and a dedicated deep-learning noise reduction algorithm

This study aimed to determine whether the whole-body bone Single Photon Emission Computed Tomography (SPECT) recording times of around 10 min, routinely provided by a high-sensitivity 360° cadmium and zinc telluride (CZT) camera, can be further reduced by a deep-learning noise reduction (DLNR) algorithm.

METHODS

DLNR was applied on whole-body images recorded after the injection of 545 ± 33 MBq of [99mTc]Tc-HDP in 19 patients (14 with bone metastasis) and reconstructed with 100%, 90%, 80%, 70%, 60%, 50%, 40%, and 30% of the original SPECT recording times.

RESULTS

Irrespective of recording time, DLNR enhanced the contrast-to-noise ratios and slightly decreased the standardized uptake values of bone lesions. Except in one markedly obese patient, the quality of DLNR processed images remained good-to-excellent down to 60% of the recording time, corresponding to around 6 min SPECT-recording.

CONCLUSION

Ultra-fast SPECT recordings of 6 min can be achieved when DLNR is applied on whole-body bone 360° CZT-SPECT.

Deep learning-based PET image denoising and reconstruction: a review

This review focuses on positron emission tomography (PET) imaging algorithms and traces the evolution of PET image reconstruction methods. First, we provide an overview of conventional PET image reconstruction methods from filtered backprojection through to recent iterative PET image reconstruction algorithms, and then review deep learning methods for PET data up to the latest innovations within three main categories. The first category involves post-processing methods for PET image denoising. The second category comprises direct image reconstruction methods that learn mappings from sinograms to the reconstructed images in an end-to-end manner. The third category comprises iterative reconstruction methods that combine conventional iterative image reconstruction with neural-network enhancement. We discuss future perspectives on PET imaging and deep learning technology.

Fully automated computational measurement of noise in positron emission tomography

OBJECTIVES

To introduce an automated computational algorithm that estimates the global noise level across the whole imaging volume of PET datasets.

METHODS

[18F]FDG PET images of 38 patients were reconstructed with simulated decreasing acquisition times (15-120 s) resulting in increasing noise levels, and with block sequential regularized expectation maximization with beta values of 450 and 600 (Q.Clear 450 and 600). One reader performed manual volume-of-interest (VOI) based noise measurements in liver and lung parenchyma and two readers graded subjective image quality as sufficient or insufficient. An automated computational noise measurement algorithm was developed and deployed on the whole imaging volume of each reconstruction, delivering a single value representing the global image noise (Global Noise Index, GNI). Manual noise measurement values and subjective image quality gradings were compared with the GNI.

RESULTS

Irrespective of the absolute noise values, there was no significant difference between the GNI and manual liver measurements in terms of the distribution of noise values (p = 0.84 for Q.Clear 450, and p = 0.51 for Q.Clear 600). The GNI showed a fair to moderately strong correlation with manual noise measurements in liver parenchyma (r = 0.6 in Q.Clear 450, r = 0.54 in Q.Clear 600, all p < 0.001), and a fair correlation with manual noise measurements in lung parenchyma (r = 0.52 in Q.Clear 450, r = 0.33 in Q.Clear 600, all p < 0.001). Classification performance of the GNI for subjective image quality was AUC 0.898 for Q.Clear 450 and 0.919 for Q.Clear 600.

CONCLUSION

An algorithm provides an accurate and meaningful estimation of the global noise level encountered in clinical PET imaging datasets.

CLINICAL RELEVANCE STATEMENT

An automated computational approach that measures the global noise level of PET imaging datasets may facilitate quality standardization and benchmarking of clinical PET imaging within and across institutions.

KEY POINTS

• Noise is an important quantitative marker that strongly impacts image quality of PET images. • An automated computational noise measurement algorithm provides an accurate and meaningful estimation of the global noise level encountered in clinical PET imaging datasets. • An automated computational approach that measures the global noise level of PET imaging datasets may facilitate quality standardization and benchmarking as well as protocol harmonization.

M3S-Net: multi-modality multi-branch multi-self-attention network with structure-promoting loss for low-dose PET/CT enhancement

Objective.PET (Positron Emission Tomography) inherently involves radiotracer injections and long scanning time, which raises concerns about the risk of radiation exposure and patient comfort. Reductions in radiotracer dosage and acquisition time can lower the potential risk and improve patient comfort, respectively, but both will also reduce photon counts and hence degrade the image quality. Therefore, it is of interest to improve the quality of low-dose PET images.Approach.A supervised multi-modality deep learning model, named M3S-Net, was proposed to generate standard-dose PET images (60 s per bed position) from low-dose ones (10 s per bed position) and the corresponding CT images. Specifically, we designed a multi-branch convolutional neural network with multi-self-attention mechanisms, which first extracted features from PET and CT images in two separate branches and then fused the features to generate the final generated PET images. Moreover, a novel multi-modality structure-promoting term was proposed in the loss function to learn the anatomical information contained in CT images.Main results.We conducted extensive numerical experiments on real clinical data collected from local hospitals. Compared with state-of-the-art methods, the proposed M3S-Net not only achieved higher objective metrics and better generated tumors, but also performed better in preserving edges and suppressing noise and artifacts.Significance.The experimental results of quantitative metrics and qualitative displays demonstrate that the proposed M3S-Net can generate high-quality PET images from low-dose ones, which are competable to standard-dose PET images. This is valuable in reducing PET acquisition time and has potential applications in dynamic PET imaging.

Low 18F-fluorodeoxyglucose dose positron emission tomography assisted by a deep-learning image-denoising technique in patients with lymphoma

Background

Patients with lymphoma receive multiple positron emission tomography/computed tomography (PET/CT) exams for monitoring of the therapeutic response. With PET imaging, a reduced level of injected fluorine-18 fluorodeoxyglucose ([18F]FDG) activity can be administered while maintaining the image quality. In this study, we investigated the efficacy of applying a deep learning (DL) denoising-technique on image quality and the quantification of metabolic parameters and Deauville score (DS) of a low [18F]FDG dose PET in patients with lymphoma.

Methods

This study retrospectively enrolled 62 patients who underwent [18F]FDG PET scans. The low-dose (LD) data were simulated by taking a 50% duration of routine-dose (RD) PET list-mode data in the reconstruction, and a U-Net-based denoising neural network was applied to improve the images of LD PET. The visual image quality score (1 = undiagnostic, 5 = excellent) and DS were assessed in all patients by nuclear radiologists. The maximum, mean, and standard deviation (SD) of the standardized uptake value (SUV) in the liver and mediastinum were measured. In addition, lesions in some patients were segmented using a fixed threshold of 2.5, and their SUV, metabolic tumor volume (MTV), and tumor lesion glycolysis (TLG) were measured. The correlation coefficient and limits of agreement between the RD and LD group were analyzed.

Results

The visual image quality of the LD group was improved compared with the RD group. The DS was similar between the RD and LD group, and the negative (DS 1-3) and positive (DS 4-5) results remained unchanged. The correlation coefficients of SUV in the liver, mediastinum, and lesions were all >0.85. The mean differences of SUVmax and SUVmean between the RD and LD groups, respectively, were 0.22 [95% confidence interval (CI): -0.19 to 0.64] and 0.02 (95% CI: -0.17 to 0.20) in the liver, 0.13 (95% CI: -0.17 to 0.42) and 0.02 (95% CI: -0.12 to 0.16) in the mediastinum, and -0.75 (95% CI: -3.42 to 1.91), and -0.13 (95% CI: -0.57 to 0.31) in lesions. The mean differences in MTV and TLG were 0.85 (95% CI: -2.27 to 3.98) and 4.06 (95% CI: -20.53 to 28.64) between the RD and LD groups.

Conclusions

The DL denoising technique enables accurate tumor assessment and quantification with LD [18F]FDG PET imaging in patients with lymphoma.

The Past, Present, and Future Role of Artificial Intelligence in Ventilation/Perfusion Scintigraphy: A Systematic Review

Ventilation-perfusion (V/Q) lung scans constitute one of the oldest nuclear medicine procedures, remain one of the few studies performed in the acute setting, and are amongst the few performed in the emergency setting. V/Q studies have witnessed a long fluctuation in adoption rates in parallel to continuous advances in image processing and computer vision techniques. This review provides an overview on the status of artificial intelligence (AI) in V/Q scintigraphy. To clearly assess the past, current, and future role of AI in V/Q scans, we conducted a systematic Ovid MEDLINE(R) literature search from 1946 to August 5, 2022 in addition to a manual search. The literature was reviewed and summarized in terms of methodologies and results for the various applications of AI to V/Q scans. The PRISMA guidelines were followed. Thirty-one publications fulfilled our search criteria and were grouped into two distinct categories: (1) disease diagnosis/detection (N = 22, 71.0%) and (2) cross-modality image translation into V/Q images (N = 9, 29.0%). Studies on disease diagnosis and detection relied heavily on shallow artificial neural networks for acute pulmonary embolism (PE) diagnosis and were primarily published between the mid-1990s and early 2000s. Recent applications almost exclusively regard image translation tasks from CT to ventilation or perfusion images with modern algorithms, such as convolutional neural networks, and were published between 2019 and 2022. AI research in V/Q scintigraphy for acute PE diagnosis in the mid-90s to early 2000s yielded promising results but has since been largely neglected and thus have yet to benefit from today's state-of-the art machine-learning techniques, such as deep neural networks. Recently, the main application of AI for V/Q has shifted towards generating synthetic ventilation and perfusion images from CT. There is therefore considerable potential to expand and modernize the use of real V/Q studies with state-of-the-art deep learning approaches, especially for workflow optimization and PE detection at both acute and chronic stages. We discuss future challenges and potential directions to compensate for the lag in this domain and enhance the value of this traditional nuclear medicine scan.

Feasibility of a deep learning algorithm to achieve the low-dose 68Ga-FAPI/the fast-scan PET images: a multicenter study

OBJECTIVES

Our work aims to study the feasibility of a deep learning algorithm to reduce the 68Ga-FAPI radiotracer injected activity and/or shorten the scanning time and to investigate its effects on image quality and lesion detection ability.

METHODS

The data of 130 patients who underwent 68Ga-FAPI positron emission tomography (PET)/CT in two centers were studied. Predicted full-dose images (DL-22%, DL-28% and DL-33%) were obtained from three groups of low-dose images using a deep learning method and compared with the standard-dose images (raw data). Injection activity for full-dose images was 2.16 ± 0.61 MBq/kg. The quality of the predicted full-dose PET images was subjectively evaluated by two nuclear physicians using a 5-point Likert scale, and objectively evaluated by the peak signal-to-noise ratio, structural similarity index and root mean square error. The maximum standardized uptake value and the mean standardized uptake value (SUVmean) were used to quantitatively analyze the four volumes of interest (the brain, liver, left lung and right lung) and all lesions, and the lesion detection rate was calculated.

RESULTS

Data showed that the DL-33% images of the two test data sets met the clinical diagnosis requirements, and the overall lesion detection rate of the two centers reached 95.9%.

CONCLUSION

Through deep learning, we demonstrated that reducing the 68Ga-FAPI injected activity and/or shortening the scanning time in PET/CT imaging was feasible. In addition, 68Ga-FAPI dose as low as 33% of the standard dose maintained acceptable image quality.

ADVANCES IN KNOWLEDGE

This is the first study of low-dose 68Ga-FAPI PET images from two centers using a deep learning algorithm.

Imaging quality of an artificial intelligence denoising algorithm: validation in 68Ga PSMA-11 PET for patients with biochemical recurrence of prostate cancer

BACKGROUND

⁶⁸ Ga-PSMA PET is the leading prostate cancer imaging technique, but the image quality remains noisy and could be further improved using an artificial intelligence-based denoising algorithm. To address this issue, we analyzed the overall quality of reprocessed images compared to standard reconstructions. We also analyzed the diagnostic performances of the different sequences and the impact of the algorithm on lesion intensity and background measures.

METHODS

We retrospectively included 30 patients with biochemical recurrence of prostate cancer who had undergone ⁶⁸ Ga-PSMA-11 PET-CT. We simulated images produced using only a quarter, half, three-quarters, or all of the acquired data material reprocessed using the SubtlePET® denoising algorithm. Three physicians with different levels of experience blindly analyzed every sequence and then used a 5-level Likert scale to assess the series. The binary criterion of lesion detectability was compared between series. We also compared lesion SUV, background uptake, and diagnostic performances of the series (sensitivity, specificity, accuracy).

RESULTS

VPFX-derived series were classified differently but better than standard reconstructions (p < 0.001) using half the data. Q.Clear series were not classified differently using half the signal. Some series were noisy but had no significant effect on lesion detectability (p > 0.05). The SubtlePET® algorithm significantly decreased lesion SUV (p < 0.005) and increased liver background (p < 0.005) and had no substantial effect on the diagnostic performance of each reader.

CONCLUSION

We show that the SubtlePET® can be used for ⁶⁸ Ga-PSMA scans using half the signal with similar image quality to Q.Clear series and superior quality to VPFX series. However, it significantly modifies quantitative measurements and should not be used for comparative examinations if standard algorithm is applied during follow-up.

Deep Learning Denoising Improves and Homogenizes Patient [18F]FDG PET Image Quality in Digital PET/CT

Given the constant pressure to increase patient throughput while respecting radiation protection, global body PET image quality (IQ) is not satisfactory in all patients. We first studied the association between IQ and other variables, in particular body habitus, on a digital PET/CT. Second, to improve and homogenize IQ, we evaluated a deep learning PET denoising solution (Subtle PET^TM) using convolutional neural networks. We analysed retrospectively in 113 patients visual IQ (by a 5-point Likert score in two readers) and semi-quantitative IQ (by the coefficient of variation in the liver, CV_liv) as well as lesion detection and quantification in native and denoised PET. In native PET, visual and semi-quantitative IQ were lower in patients with larger body habitus (p < 0.0001 for both) and in men vs. women (p ≤ 0.03 for CV_liv). After PET denoising, visual IQ scores increased and became more homogeneous between patients (4.8 ± 0.3 in denoised vs. 3.6 ± 0.6 in native PET; p < 0.0001). CV_liv were lower in denoised PET than in native PET, 6.9 ± 0.9% vs. 12.2 ± 1.6%; p < 0.0001. The slope calculated by linear regression of CV_liv according to weight was significantly lower in denoised than in native PET (p = 0.0002), demonstrating more uniform CV_liv. Lesion concordance rate between both PET series was 369/371 (99.5%), with two lesions exclusively detected in native PET. SUV_max and SUV_peak of up to the five most intense native PET lesions per patient were lower in denoised PET (p < 0.001), with an average relative bias of -7.7% and -2.8%, respectively. DL-based PET denoising by Subtle PET^TM allowed [¹⁸F]FDG PET global image quality to be improved and homogenized, while maintaining satisfactory lesion detection and quantification. DL-based denoising may render body habitus adaptive PET protocols unnecessary, and pave the way for the improvement and homogenization of PET modalities.

Artificial intelligence-based 68Ga-DOTATOC PET denoising for optimizing 68Ge/68Ga generator use throughout its lifetime

Introduction

The yield per elution of a 68Ge/68Ga generator decreases during its lifespan. This affects the number of patients injected per elution or the injected dose per patient, thereby negatively affecting the cost of examinations and the quality of PET images due to increased image noise. We aimed to investigate whether AI-based PET denoising can offset this decrease in image quality parameters.

Methods

All patients addressed to our PET unit for a 68Ga-DOTATOC PET/CT from April 2020 to February 2021 were enrolled. Forty-four patients underwent their PET scans according to Protocol_FixedDose (150 MBq) and 32 according to Protocol_WeightDose (1.5 MBq/kg). Protocol_WeightDose examinations were processed using the Subtle PET software (Protocol_WeightDoseAI). Liver and vascular SUV mean were recorded as well as SUVmax, SUVmean and metabolic tumour volume (MTV) of the most intense tumoural lesion and its background SUVmean. Liver and vascular coefficients of variation (CV), tumour-to-background and tumour-to-liver ratios were calculated.

Results

The mean injected dose of 2.1 (0.4) MBq/kg per patient was significantly higher in the Protocol_FixedDose group as compared to 1.5 (0.1) MBq/kg for the Protocol_WeightDose group. Protocol_WeightDose led to noisier images than Protocol_FixedDose with higher CVs for liver (15.57% ± 4.32 vs. 13.04% ± 3.51, p = 0.018) and blood-pool (28.67% ± 8.65 vs. 22.25% ± 10.37, p = 0.0003). Protocol_WeightDoseAI led to less noisy images than Protocol_WeightDose with lower liver CVs (11.42% ± 3.05 vs. 15.57% ± 4.32, p &lt; 0.0001) and vascular CVs (16.62% ± 6.40 vs. 28.67% ± 8.65, p &lt; 0.0001). Tumour-to-background and tumour-to-liver ratios were lower for protocol_WeightDoseAI: 6.78 ± 3.49 vs. 7.57 ± 4.73 (p = 0.01) and 5.96 ± 5.43 vs. 6.77 ± 6.19 (p &lt; 0.0001), respectively. MTVs were higher after denoising whereas tumour SUVmax were lower: the mean% differences in MTV and SUVmax were + 11.14% (95% CI = 4.84–17.43) and −3.92% (95% CI = −6.25 to −1.59).

Conclusion

The degradation of PET image quality due to a reduction in injected dose at the end of the 68Ge/68Ga generator lifespan can be effectively counterbalanced by using AI-based PET denoising.

Feasibility of Dedicated Breast Positron Emission Tomography Image Denoising Using a Residual Neural Network

Objectives

This study aimed to create a deep learning (DL)-based denoising model using a residual neural network (Res-Net) trained to reduce noise in ring-type dedicated breast positron emission tomography (dbPET) images acquired in about half the emission time, and to evaluate the feasibility and the effectiveness of the model in terms of its noise reduction performance and preservation of quantitative values compared to conventional post-image filtering techniques.

Methods

Low-count (LC) and full-count (FC) PET images with acquisition durations of 3 and 7 minutes, respectively, were reconstructed. A Res-Net was trained to create a noise reduction model using fifteen patients' data. The inputs to the network were LC images and its outputs were denoised PET (LC + DL) images, which should resemble FC images. To evaluate the LC + DL images, Gaussian and non-local mean (NLM) filters were applied to the LC images (LC + Gaussian and LC + NLM, respectively). To create reference images, a Gaussian filter was applied to the FC images (FC + Gaussian). The usefulness of our denoising model was objectively and visually evaluated using test data set of thirteen patients. The coefficient of variation (CV) of background fibroglandular tissue or fat tissue were measured to evaluate the performance of the noise reduction. The SUV_max and SUV_peak of lesions were also measured. The agreement of the SUV measurements was evaluated by Bland-Altman plots.

Results

The CV of background fibroglandular tissue in the LC + DL images was significantly lower (9.10±2.76) than the CVs in the LC (13.60± 3.66) and LC + Gaussian images (11.51± 3.56). No significant difference was observed in both SUV_max and SUV_peak of lesions between LC + DL and reference images. For the visual assessment, the smoothness rating for the LC + DL images was significantly better than that for the other images except for the reference images.

Conclusion

Our model reduced the noise in dbPET images acquired in about half the emission time while preserving quantitative values of lesions. This study demonstrates that machine learning is feasible and potentially performs better than conventional post-image filtering in dbPET denoising.