Sensitivity of an AI method for [18F]FDG PET/CT outcome prediction of Diffuse large B-cell lymphoma patients to image reconstruction protocols.

doi:10.21203/rs.3.rs-3128710/v1

Download PDF

Research Article

Sensitivity of an AI method for [18F]FDG PET/CT outcome prediction of Diffuse large B-cell lymphoma patients to image reconstruction protocols.

https://doi.org/10.21203/rs.3.rs-3128710/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 28 Sep, 2023

Read the published version in EJNMMI Research →

You are reading this latest preprint version

Background

Convolutional Neural Networks (CNN), applied to baseline [¹⁸F]-FDG PET/CT maximum intensity projections (MIPs) show potential for treatment outcome prediction in diffuse large B-cell lymphoma (DLBCL). The aim of this study is to investigate the robustness of CNN predictions to different image reconstruction protocols, such as the European Association of Nuclear Medicine Research Ltd. (EARL) harmonization standards 1 and 2. Baseline [¹⁸F]FDG PET/CT scans were collected from 20 DLBCL patients. EARL1, EARL2 and high-resolution or HR (point spread function with pixel spacing of 2mm) protocols were applied per scan, generating 3 images with different image qualities. An in-house developed CNN predicted the probability of tumor progression within 2 years for each patient for the 3 reconstructions. Image-based transformation was applied by blurring EARL2 and HR images to generate EARL1 compliant images using a Gaussian filter of 5 and 7 mm, respectively. The difference in probabilities per patient was then calculated between both EARL2 and HR with respect to EARL1 (delta probabilities or ΔP). We compared these to the probabilities obtained after aligning the data with ComBat.

Results

CNN probabilities were found to be sensitive to different reconstruction protocols (EARL2 ΔP: median = 0.09, interquartile range (IQR) = [0.06, 0.10] and HR ΔP: median = 0.1, IQR = [0.08, 0.16]). Moreover, higher resolution images (EARL2 and HR) led to higher probability values. After image-based and ComBat transformation, an improved agreement of CNN probabilities among reconstructions was found for all patients. This agreement was slightly better after image-based transformation (transformed EARL2 ΔP: median = 0.022, IQR = [0.01, 0.02] and transformed HR ΔP: median = 0.029, IQR= [0.01, 0.03]).

Conclusion

Our CNN-based outcome predictions are affected by the applied reconstruction protocols, yet in a predictable manner. Image-based harmonization is a suitable approach to harmonize CNN predictions across image reconstruction protocols.

diffuse large B-cell lymphoma

PET

convolutional neural networks

reconstruction

Diffuse large B cell lymphoma (DLBCL) is the most common subtype of non-Hodgkin lymphoma, accounting for 30 to 40% of all cases (1). [¹⁸F]-Fluorodeoxyglucose ([¹⁸F]FDG) positron emission tomography (PET) in combination with computed tomography (CT) imaging is widely used for diagnosis, staging, prognosis, prediction and response monitoring in DLBCL patients (2). Different metrics such as metabolic tumor volume (MTV), standard uptake value (SUV) or the maximum distance between two lesions (Dmax) are extracted from these images which provide insight into the tumor characteristics. Baseline MTV has proven to be a strong prognostic factor for tumor progression in lymphoma, alongside with disease dissemination features (3–5). In order to extract these features, the tumor needs to be delineated. This task is time consuming, user dependent and suffers from inter and intra reader variability. These limitations could be overcome with the help of artificial Intelligence (AI). AI could offer a solution in the form of an automated method for disease progression and treatment monitoring (6). Convolutional neural networks (CNN) are currently being investigated for this purpose and were found to have potential for treatment outcome prediction in DLBCL (7, 8).

Technical aspects of PET imaging, such as image acquisition and reconstruction settings, should be taken into account when analyzing PET scans as these may have an impact on the derived metrics (9, 10). ComBat harmonization can be used to reduce the variability generated by these technical aspects (11). To accomplish this, ComBat standardizes the means and the variances across the batches of data derived from different scanners/protocols (12). In our previous study, ComBat was partially able to reduce reconstruction-dependent MTV variability (9). Kaalep et al (13), addressed this issue by altering the images instead of the data. They applied Gaussian filters to scans obtained from European Association of Nuclear Medicine Research Ltd. (EARL) harmonization standards 2. These filters were used to ‘blur’ the images, mimicking the resolution of the EARL1 scans.

As with MTV, PET imaging aspects may also have an impact on PET-based CNN performance. A recent study showed that PET images derived from block sequential regularized expectation maximization reconstruction yielded a better CNN performance than ordered subset expectation maximization (OSEM) reconstruction for the detection of pulmonary lesions (14). We recently developed a CNN to predict the probability of 2-year time to progression (TTP) using maximum intensity projections (MIP) of [¹⁸F]FDG PET/CT baseline scans of DLBCL patients (8). The CNN achieved an area under the curve of 0.72 in the internal validation and of 0.74 in the external validation. The model was, however, trained on retrospective data from the HOVON84 trial (15). This data was predominantly collected on older generation of PET/CT systems. When using our CNN with data collected on state-of-the-art systems this will possibly imply PET images with higher spatial resolution.

Little is known about how PET-based classification CNNs are affected by differences in image quality, for example, when sites use PET images compliant with EARL standards 1 and 2 or images reconstructed with locally preferred clinical protocol. Therefore, in this study, we assessed the robustness of outcome predictions provided by our recently developed CNN to different reconstruction protocols. Furthermore, we assessed the ability of image-based transformations using Gaussian filters to mitigate the predicted probabilities variability caused by the use of different reconstruction protocols and compared it to ComBat transformation results. The main difference between these two transformations is that the image-based transformation is applied to the scans during the preprocessing stage, unlike ComBat, which is a linear transformation applied directly onto the CNN probabilities.

Study Population

For the analysis, we used baseline [¹⁸F]FDG PET/CT scans from 20 DLBCL patients. From these, 13 patients had been scanned at Amsterdam UMC and were retrospectively obtained from medical records, with a waiver for informed consent from the Medical Ethics Review Committee of Amsterdam UMC, location VUmc. This study was registered as IRB2018.029. The other 7 patients were recruited and scanned at the outpatient clinic of the department of Hematology of the Amsterdam UMC, location VUmc and the outpatient clinic of the department of Hematology of the Amstelland Hospital in Amstelveen (IRB2019.278). Patients included in these trials required to be 18 years or older and have at least one tumor with a diameter of 3 cm or more. Patients with metal implants, multiple malignancies, who had undergone chemotherapy in the past 4 weeks or who were pregnant/lactating were excluded from the trials.

Image acquisition

Patients scans were performed on two EARL-accredited Philips scanners, Ingenuity TF PET/CT and Vereos PET/CT, with BLOB-OS-TF reconstruction method and an [¹⁸F]FDG uptake time of 60 minutes. The mean injected activity was 264.12 megabecquerels (MBq).

Image Processing

Three different reconstructions protocols were used to derive the scans: following EARL1 standards (EARL1 reconstruction), following EARL2 standards (EARL2 reconstruction) and a third reconstruction which followed locally clinically preferred protocols (High Resolution or HR reconstruction). EARL2 standards were established by introducing a resolution modelling algorithm, point spread function (PSF), to the initial EARL1 standards (13). The use of PFS improves image resolution and contrast (16). The main difference between the HR and the EARL reconstructions is that HR introduced a pixel spacing of 2mm instead of 4mm to achieve a higher spatial resolution. Technical details about these 3 protocols can be found in our previous study (9).

For every scan and reconstruction, MIPs were generated using an in-house developed CNN preprocessing tool. This tool converted scans into coronal and sagittal MIPs with size 275 x 200 x 1 and pixel size of 4 x 4 mm. Furthermore, the brain was removed from each of the MIPs with the aim of providing greater consistency across the dataset since not all scans fully included the head. More details about this process can be found in our previous study (8). An example of MIPs for each reconstruction is given in Fig. 1 (A), illustrating the difference in spatial resolution between the various reconstructions

Quality Control

Quality control (QC) of baseline [¹⁸F]FDG PET/CT scans was performed following the criteria described by the EANM guidelines. The liver mean standardized uptake value (SUVmean) should range between 1.3 and 3.0 and the plasma glucose should not surpass 11 mmol/L (2).

Harmonization

Two different harmonization techniques were used in this study to align the CNN predicted probabilities between the three different reconstructions: EARL1, EARL2, and HR. We implemented an image-based transformation to change the resolution of our EARL2 and HR images to resemble EARL1. Moreover, we implemented a data transformation using ComBat to align the CNN probabilities of the EARL2 and HR images to those of EARL1.

Image-based harmonization. The ACCURATE software was used to perform the image-based transformation of the EARL2 and HR images in order to match image qualities (in particular spatial resolution) with that of EARL1 reconstructions (17). A Gaussian filter (full width at half maximum, FWHM) of 5 mm was applied to the EARL2 images and of 7 mm to the HR images to match the spatial resolution to that of EARL1 reconstruction (18).

ComBat harmonization. ComBat was applied to the CNN probabilities yielded by the non-transformed EARL1, EARL2 and HR scans. We used ComBat to align the mean and the standard deviation of the probabilities obtained from HR and EARL2 MIPs to those of EARL1. Combat was applied using R version 4.0.5 on the code provided by Fortin et al. (19).

Convolutional Neural Network

A previously developed CNN to predict the probability of 2 year TTP in DLBCL patients from MIP images is utilized in this study (8). The CNN was trained following a 5-fold cross validation and further validated on an external dataset. The CNN consists of two branches which analyze both coronal and sagittal MIPs through a series of convolutional and max pooling layer and, concatenates with a final fully connected layer. Technical details about the cross validation and the design of the CNN can be found in (8). The output is a binary prediction given by the probability of TTP longer than 2 years (TTP0) or TTP shorter than 2 years (TTP1), where TTP1 indicates an increased risk of tumour progression or recurrence for the patient. TTP0 may indicate absence of tumour progression or absence of recurrence. In this study we used the CNN to predict TTP1 from the EARL1, EARL2 and HR reconstructions and assessed the impact on those probabilities after image-based and ComBat transformations.

Statistical Analysis

A probability per patient is obtained for each of the three different reconstructions. A non-parametrical statistical hypothesis test, Wilcoxon signed-rank test, was used to compare probabilities between the reconstructions before and after transformation, (image-based and ComBat). The probability difference (delta probabilities or ∆P) was calculated for both EARL2 and HR with respect to EARL1. For EARL2 and HR reconstructions, the median and the interquartile range (IQR) of ∆P was calculated for both scenarios; before and after transformations. Moreover, to estimate the strength of the association between probabilities we used regression analyses and Bland-Altman plot.

Patients characteristics

There were a total of 20 patients included in this study. A summary of patients characteristics relevant to the study are given in Table 1. All of the scans complied with the QC requirements.

Table 1

Patients Characteristics.
Patients Characteristics
Sex (N)
Male (%) Female (%)	5 (25%) 15 (75%)
Weight (kg)
Mean (min - max)	77.35 (54–103)
Height (cm)
Mean (min - max)	176.4 (160–190)
Injected activity (MBq)
Mean (min - max)	264.12 (164.87–367.48)

CNN probabilities

Wilcoxon signed-rank test was used to compare paired probabilities with EARL1 as the reference reconstruction. Statistical differences were found before transformation between EARL1 and HR (p-value lower than 0.05). These differences were considerably decreased when comparing to probabilities after both image and Combat transformation. The resulting p-values per comparison can be found in Table 2.

EARL1 probabilities were generally lower compared to EARL2 and HR probabilities (EARL1 probabilities: median = 0.47 and IQR = [0.38, 0.61], EARL2: median = 0.56 and IQR = [0.46, 0.7], HR: median = 0.64 and IQR = [0.53, 0.74]). An example of the 3 different reconstructions with their corresponding CNN probabilities can be found in Fig. 1 (A). The differences were greater between EARL1 and HR than between EARL1 and EARL2 probabilities (Table 3). This suggests that an increase in the image resolution leads to an increase in the probability values for the same patient. Figure 2 illustrates the relationship between the probability values generated from different reconstructions is mostly linear, thus, there is a predictable change in the probabilities.

Image-based transformation

After image-based transformation, the overall differences in probabilities (ΔP) were considerably decreased and the CNN probabilities showed an improved agreement between reconstructions (Table 3). This is illustrated in Fig. 2 where the regression line of the transformed probabilities values is closer to the line of identity compared to the regression line of the original values. The Bland-Altman plots in Fig. 3 (A-D) show an improve agreement of the probabilities after image transformation. In Fig. 1 (B) the same example as in Fig. 1 (A) is shown with the MIPs after image-based transformation and their corresponding probabilities.

ComBat transformation

A similar trend was observed for the ComBat transformed probabilities. The regression line for the Combat-transformed values is shown in Fig. 2, together with the regression line for the image-based transformed values. An improved agreement of the probabilities after ComBat transformation is shown in Fig. 3 (A-D).

The ΔP was very similar for both image-based and ComBat transformations. However, ComBat resulted in a slightly worsened alignment of the probabilities for the HR reconstruction (Fig. 1 (B)), where there is a greater variability given by the ΔP IQ and also observable in Fig. 2 (B).

Table 2

P-values from non-parametric Wilcoxon signed-rank test for comparing probabilities between reconstruction with EARL1 as reference. Statistical differences shown with an asterisk (p-value < 0.05).
	EARL2-EARL1: P-value	HR-EARL1: P-value
Original values	0.055	0.012*
Image-based transformation	0.41	0.79
ComBat transformation	0.88	0.85

Table 3

Overall differences in probabilities (ΔP) between reconstructions with EARL1 as reference.
	EARL2 ∆P: Median (IQ)	HR ∆P : Median (IQ)
Original values	0.09 (0.06, 0.10)	0.10 (0.08, 0.16)
Image-based transformation	0.02 (0.01, 0.03)	0.03 (0.01, 0.03)
ComBat transformation	0.02 (0.01, 0.03)	0.04 (0.03, 0.06)

The aim of this study was to assess the robustness of a CNN model to PET images derived from different reconstruction protocols. The CNN is applied to MIP images generated from [¹⁸F]FDG PET/CT baseline scans and used to predict the probability of 2 year TTP in DLBCL patients. This model has been internally and externally validated in a previous study (8). Although CNNs show potential to improve the current state of prognosis in lymphoma (7, 20, 21), there is still an acknowledged need to develop robust and reproducible prognostic markers to eventually attain their clinical implementation (22–25).

In this study, we found considerable differences between the resulting probabilities depending on the reconstruction protocol. For EARL1 images, the probabilities tend to be lower compared to the EARL2 and HR images. The probabilities generated from the HR images, which have the highest resolution in this study, were generally higher across the 20 patients. Overall, we observed that the CNN probabilities were affected by the image resolution in a predictable manner (higher probability values at higher resolution images, i.e. EARL2 and HR). The impact of reconstruction protocols on PET-derived measurements has been assessed in previous studies. SUVmax, SUVpeak, MTV and multiple textural features have all been found to be sensitive to changes in reconstruction protocols (9, 16, 26, 27). In our previous study, we mitigated the variability among MTV values from images reconstructed with different protocols by harmonizing the data using ComBat (9). Herein, we assessed the implementation of ComBat for the harmonization of the CNN probabilities and compared it to an image-based transformation. An improved agreement between the CNN probabilities was achieved by both transformations. ComBat applies a linear transformation to the data, which seems to agree with the relationship seen among the CNN probabilities between the different reconstructions. Nevertheless, the image-based transformation was best in aligning the values for the HR reconstruction. These results indicate that transforming the images beforehand could aid in the development of a robust and reliable metric for prognosis and, moreover, facilitate the implementation of CNN models for treatment outcome assessment. When using differently reconstructed images without any transformation, large differences in estimated probabilities may occur (Fig. 1) that can potentially affect clinical decision making. Therefore, reliable predictions using the CNN can only be made when information on the applied acquisition and reconstruction settings are available. In this way the appropriate image transformation can be applied such that the spatial resolution required by the CNN (i.e. conform EARL1) is obtained assuring reproducible estimation of probabilities.

A possible limitation of our study is the small number of subjects. However, by performing three different reconstructions protocols for each scan, we were able to directly compare the impact of image quality on CNN based predictions and allowing us to perform a head-to-head comparison of these probabilities. We observed very clear relationships among the differently obtained probabilities. Moreover, despite the small number, we believe that we convincingly showed that image transformations can be successfully applied to mitigate reconstruction effects. Therefore adding more subjects will not change our main findings and conclusions. Another limitation could be that transformations can only be applied when the actual image spatial resolution is known. In most of the cases this can be derived from the DICOM header information, providing details on voxel size and image reconstruction settings. Yet, when using anonymized images, this information is sometimes missing and, therefore, future work will focus on automatically assessing the effective spatial resolution of reconstructed PET data, e.g. using the CNN approach proposed by Pfaehler et al (28).

The CNN predicted probabilities are affected by the applied reconstruction protocol, yet in a predictable manner; higher resolution images (i.e. EARL2 and HR) resulted in higher probability values. After ComBat and image-based transformation, the EARL2 and HR probabilities were closely aligned with EARL1 probabilities. The image-based transformation mitigated the differences slightly better. These findings suggest that image-based transformation is a suitable approach for harmonization of CNN predictions across image reconstruction protocols.

Diffuse Large B-Cell Lymphoma (DLBCL)

Positron Emission Tomography (PET)

Computed Tomography (CT)

Metabolic Tumor Volume (MTV)

[¹⁸F]-Fluorodeoxyglucose ([¹⁸F]FDG)

Artificial Intelligence (AI)

Time to progression (TTP)

Point Spread Function (PSF)

Standardized Uptake Value (SUV)

High Resolution (HR)

Quality Control (QC)

Convolutional Neural Networks (CNN)

Interquartile Range (IQR)

European Association of Nuclear Medicine (EANM) EANM Research Ltd. (EARL)

Funding: Hanarth Fonds Fund and the Dutch Cancer Society (#VU-2018-11648)

Acknowledgements. This work was financially supported by the Hanarth Fonds Fund and the Dutch Cancer Society. The sponsor had no role in gathering, analyzing or interpreting the data. The authors thank all the patients who participated in the trial.

Authors contributions. All authors contributed to the design of the study. G.J.C.Z. , S.P., L.S. and R.B. were responsible for acquiring and collecting the data. S.P., S.E.W and M.C.F. performed the data analysis. M.C.F. completed the first draft of the manuscript. B.M.d.V, M.W.H, S.E.W, J.J.E., S.S.V.G., R.B., G.J.C.Z. and J.M.Z. reviewed and approved the manuscript. All authors read and approved the final manuscript.

Funding. This work is financially supported by the Hanarth Fonds Fund and the Dutch Cancer Society (#VU-2018-11648).

Availability of data and materials. The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Code availability. Not applicable

Ethics approval and consent to participate. This study was conducted in accordance with the ethical standards as laid down in the Declaration of Helsinki and its later amendments. Ethics approval was obtained by the Medical Ethics Review Committee of Amsterdam UMC, location VUmc (IRB2018.029) and the department of Hematology of the Amstelland Hospital in Amstelveen (IRB2019.278). Informed consent was waived given the retrospective nature of the study.

Consent for publication. Not applicable

Competing interests. This work was financially supported by the Hanarth Fonds Fund and the Dutch Cancer Society (#VU-2018-11648). M.C.F., S.S.V.G, J.J.E., B.M.d.V.,S.E.W., G.J.C.Z., S.P. and R.B. declare no competing financial interests. J.M.Z. received research funding from Roche and received honoraria for advisory boards from Takeda, Gilead, BMS and Roche. No other potential conflicts of interest relevant to this article exist.

Crump M, Neelapu SS, Farooq U, Van Den Neste E, Kuruvilla J, Westin J, et al. Outcomes in refractory diffuse large B-cell lymphoma: results from the international SCHOLAR-1 study. Blood. 2017;130(16):1800–8. https://doi.org/10.1182/blood-2017-03-769620.
Boellaard R, Delgado-Bolton R, Oyen WJ, Giammarile F, Tatsch K, Eschner W, et al. FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0. Eur J Nucl Med Mol Imaging. 2015;42(2):328–54. https://doi.org/10.1007/s00259-014-2961-x.
Eertink JJ, van de Brug T, Wiegers SE, Zwezerijnen GJC, Pfaehler EAG, Lugtenburg PJ, et al. (18)F-FDG PET baseline radiomics features improve the prediction of treatment outcome in diffuse large B-cell lymphoma. Eur J Nucl Med Mol Imaging. 2022;49(3):932–42. https://doi.org/10.1007/s00259-021-05480-3.
Cottereau AS, Nioche C, Dirand AS, Clerc J, Morschhauser F, Casasnovas O, et al. (18)F-FDG PET Dissemination Features in Diffuse Large B-Cell Lymphoma Are Predictive of Outcome. J Nucl Med. 2020;61(1):40–5. https://doi.org/10.2967/jnumed.119.229450.
Schmitz C, Huttmann A, Muller SP, Hanoun M, Boellaard R, Brinkmann M, et al. Dynamic risk assessment based on positron emission tomography scanning in diffuse large B-cell lymphoma: Post-hoc analysis from the PETAL trial. Eur J Cancer. 2020;124:25–36. https://doi.org/10.1016/j.ejca.2019.09.027.
Heagerty PJ, Lumley T, Pepe MS. Time-dependent ROC curves for censored survival data and a diagnostic marker. Biometrics. 2000;56(2):337–. https://doi.org/10.1111/j.0006-341x.2000.00337.x. 44.
Liu P, Zhang M, Gao X, Li B, Zheng G. Joint Lymphoma Lesion Segmentation and Prognosis Prediction From Baseline FDG-PET Images via Multitask Convolutional Neural Networks. IEEE Access. 2022;10:81612–23. https://doi.org/10.1109/access.2022.3195906.
Ferrández MC, Golla SSV, Eertink JJ, de Vries BM, Lugtenburg PJ, Wiegers SE et al. An artificial intelligence method using 18F-FDG PET maximum intensity projections to predict 2-year time-to-progression in diffuse large B-cell lymphoma patients. 2023. https://doi.org/10.21203/rs.3.rs-2761494/v1.
Ferrandez MC, Eertink JJ, Golla SSV, Wiegers SE, Zwezerijnen GJC, Pieplenbosch S, et al. Combatting the effect of image reconstruction settings on lymphoma [(18)F]FDG PET metabolic tumor volume assessment using various segmentation methods. EJNMMI Res. 2022;12(1):44. https://doi.org/10.1186/s13550-022-00916-9.
Zwezerijnen GJC, Eertink JJ, Ferrandez MC, Wiegers SE, Burggraaff CN, Lugtenburg PJ, et al. Reproducibility of [18F]FDG PET/CT liver SUV as reference or normalisation factor. Eur J Nucl Med Mol Imaging. 2023;50(2):486–93. https://doi.org/10.1007/s00259-022-05977-5.
Orlhac F, Eertink JJ, Cottereau AS, Zijlstra JM, Thieblemont C, Meignan M, et al. A guide to comBat harmonization of imaging biomarkers in multicenter studies. J Nucl Med. 2022;63(2):172–9. https://doi.org/10.2967/jnumed.121.262464.
Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics. 2007;8(1):118–27. https://doi.org/10.1093/biostatistics/kxj037.
Kaalep A, Burggraaff CN, Pieplenbosch S, Verwer EE, Sera T, Zijlstra J, et al. Quantitative implications of the updated EARL 2019 PET-CT performance standards. EJNMMI Phys. 2019;6(1):28. https://doi.org/10.1186/s40658-019-0257-8.
Schwyzer M, Martini K, Benz DC, Burger IA, Ferraro DA, Kudura K, et al. Artificial intelligence for detecting small FDG-positive lung nodules in digital PET/CT: impact of image reconstructions on diagnostic performance. Eur Radiol. 2020;30(4):2031–40. https://doi.org/10.1007/s00330-019-06498-w.
Lugtenburg PJ, de Nully Brown P, van der Holt B, D'Amore FA, Koene HR, de Jongh E, et al. Rituximab-CHOP With Early Rituximab Intensification for Diffuse Large B-Cell Lymphoma: A Randomized Phase III Trial of the HOVON and the Nordic Lymphoma Group (HOVON-84). J Clin Oncol. 2020;38(29):3377–87. https://doi.org/10.1200/JCO.19.03418.
Rahmim A, Qi J, Sossi V. Resolution modeling in PET imaging: theory, practice, benefits, and pitfalls. Med Phys. 2013;40(6):064301. https://doi.org/10.1118/1.4800806.
Boellaard R. Quantitative oncology molecular analysis suite: ACCURATE. J Nucl Med. 2018;59:1753.
Jha AK, Bradshaw TJ, Buvat I, Hatt M, Kc P, Liu C, et al. Nuclear Medicine and Artificial Intelligence: Best Practices for Evaluation (the RELAINCE guidelines). J Nucl Med. 2022. https://doi.org/10.2967/jnumed.121.263239.
Fortin JP, Parker D, Tunc B, Watanabe T, Elliott MA, Ruparel K, et al. Harmonization of multi-site diffusion tensor imaging data. NeuroImage. 2017;161:149–70. https://doi.org/10.1016/j.neuroimage.2017.08.047.
Blanc-Durand P, Jegou S, Kanoun S, Berriolo-Riedinger A, Bodet-Milin C, Kraeber-Bodere F, et al. Fully automatic segmentation of diffuse large B cell lymphoma lesions on 3D FDG-PET/CT for total metabolic tumour volume prediction using a convolutional neural network. Eur J Nucl Med Mol Imaging. 2021;48(5):1362–70. https://doi.org/10.1007/s00259-020-05080-7.
Girum KB, Rebaud L, Cottereau AS, Meignan M, Clerc J, Vercellino L, et al. (18)F-FDG PET maximum intensity projections and artificial intelligence: a win-win combination to easily measure prognostic biomarkers in DLBCL patients. J Nucl Med. 2022. https://doi.org/10.2967/jnumed.121.263501.
Barrington SF, Meignan M. Time to prepare for risk adaptation in lymphoma by standardizing measurement of metabolic tumor burden. J Nucl Med. 2019;60(8):1096–102. https://doi.org/10.2967/jnumed.119.227249.
Ronald Boellard NCK, Otto S, Hoekstra, Adriaan A, Lammertsma. Effects of noise, image resolution, and ROI definition on the accuracy of standard uptake values: a simulation study. J Nucle Med. 2004;24.
Mikhaeel NG, Smith D, Dunn JT, Phillips M, Moller H, Fields PA, et al. Combination of baseline metabolic tumour volume and early response on PET/CT improves progression-free survival prediction in DLBCL. Eur J Nucl Med Mol Imaging. 2016;43(7):1209–19. https://doi.org/10.1007/s00259-016-3315-7.
Ilyas H, Mikhaeel NG, Dunn JT, Rahman F, Moller H, Smith D, et al. Defining the optimal method for measuring baseline metabolic tumour volume in diffuse large B cell lymphoma. Eur J Nucl Med Mol Imaging. 2018;45(7):1142–54. https://doi.org/10.1007/s00259-018-3953-z.
Orlhac F, Boughdad S, Philippe C, Stalla-Bourdillon H, Nioche C, Champion L, et al. A Postreconstruction harmonization method for multicenter radiomic studies in PET. J Nucl Med. 2018;59(8):1321–8. https://doi.org/10.2967/jnumed.117.199935.
Sheikhbahaei SM, Wray C, Rahmim R, Lodge A, Subramaniam MA. Impact of point spread function reconstruction on quantitative 18F-FDG-PET/CT imaging parameters and inter-reader reproducibility in solid tumors. Nucl Med Commun. 2016;37(3):288–96.
Pfaehler E, Euba D, Rinscheid A, Hoekstra OS, Zijlstra J, van Sluis J, et al. Convolutional neural networks for automatic image quality control and EARL compliance of PET images. EJNMMI Phys. 2022;9(1):53. https://doi.org/10.1186/s40658-022-00468-w.

Download PDF

Journal Publication

published 28 Sep, 2023

Read the published version in EJNMMI Research →

Editorial decision: Major Revision
21 Jul, 2023
Reviewers agreed at journal
05 Jul, 2023
Reviewers invited by journal
05 Jul, 2023
Editor assigned by journal
05 Jul, 2023
First submitted to journal
05 Jul, 2023

You are reading this latest preprint version

Sensitivity of an AI method for [18F]FDG PET/CT outcome prediction of Diffuse large B-cell lymphoma patients to image reconstruction protocols.

Status:

Journal Publication

Version 1

Abstract

Background

Results

Conclusion

Figures

BACKGROUND

METHODS

Study Population

Image acquisition

Image Processing

Quality Control

Harmonization

Convolutional Neural Network

Statistical Analysis

RESULTS

Patients characteristics

CNN probabilities

Image-based transformation

ComBat transformation

DISCUSSION

CONCLUSION

Abbreviations

Declarations

References

Status:

Journal Publication

Version 1