Skip to main content
EJNMMI Research
Download PDF
  • Original research
  • Open access
  • Published:

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

EJNMMI Research volume 14, Article number: 72 (2024) Cite this article

Abstract

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.

Background

Cancer stands as the second leading cause of death worldwide [1]. Significant therapeutic advancements have emerged in the field of oncology, especially in the past two decades [2]. Metabolic imaging, such as [18F]fluorodeoxyglucose ([18F]FDG) positron emission tomography (PET), plays a key role in the comprehensive assessment of treatment response in both clinical practice and trials [3,4,5]. Its utilization facilitates timely treatment adjustments and helps mitigate the continuation of ineffective, often expensive, and potentially toxic treatments.

In the realm of [18F]FDG PET, several classification systems have been proposed to evaluate treatment responses, recently addressing in particular immunotherapy for solid tumours [6]. The most commonly used PET response classification systems for conventional solid cancer therapy assessment are “EORTC” for “European Organisation for Research and Treatment of Cancer” [7] and more recent “PERCIST” for “PET evaluation response criteria in solid tumours” [8, 9], introducing moderate modifications compared to EORTC, and excluding PET studies considered not to be reliably assessable for response evaluation.

Notably, EORTC and PERCIST criteria yield very similar classifications in several studies [10].

Metabolic treatment response evaluation (as in [18F]FDG PET imaging) compared to anatomical assessments has been shown to be of higher and more rapid predictive value of (pathologic) response and survival outcomes in many solid cancers [9, 11,12,13,14,15]. However metabolic and anatomical information can be complementary; and metabolic signal accuracy is reduced in (small) tumours with low [18F]FDG uptake.

Both EORTC and PERCIST PET response assessments primarily hinge on the detection of new malignant [18F]FDG avid lesions in follow-up PET and on variations in lesion standardized uptake values (SUV; generally normalised on body weight; or on lean body mass (SUL)), SUVmax for EORTC and SULpeak for PERCIST. Additionally, for PERCIST, the evolution in total lesion glycolysis (TLG) may be considered, although the most adequate and discriminative measurement methodology is yet to be established.

Various technical, biological and physical factors are known to influence SUV values, thereby introducing inherent quantitative variability between two PET examinations [16,17,18]. For instance, changes in PET reconstruction parameters have been shown to affect PET response classification in a subset of patients [19, 20], highlighting the need to develop and promote standardisation and harmonisation methods for PET quantification [21].

The integration of artificial intelligence and more specifically deep learning is increasing in PET imaging [22]. Subtle PET™, a commercially available PET denoising software [23] based on convolutional neural networks, is now routinely employed in several institutions allowing either faster acquisitions [24] or injected activity reduction [25, 26].

A slight impact of this AI-based PET denoising software on lesion detection and on quantitation of lesions and of reference organs in [18F]FDG PET has been demonstrated in previous studies [24, 27, 28]. Therefore, as a next step, this study aimed to evaluate the influence of this AI PET denoising on EORTC and PERCIST treatment response classifications when comparing a standard PET to an AI denoised half-duration PET during follow-up.

Methods

Patient inclusion

Patients addressed for baseline or follow-up [18F]FDG PET/CT were prospectively included between January and February 2021 (for the purpose of a first Subtle PET™ evaluation study and of the current second study) and between August 2022 and January 2023 (only for the purpose of the current study). The only exclusion criterion was a specific acquisition protocol involving a longer acquisition time per bed position on head and neck or liver areas.

The study was approved by the institutional review ethics committee board at the François Baclesse Comprehensive Cancer Centre and registered with the French Health Data Hub under reference F20220712121936. It was conducted in compliance with the French Research Standard MR-004 (Compliance commitment to MR-004 for the Centre François Baclesse). All patients received information and none expressed opposition to the use of their data.

PET imaging protocol

“Standard” routine [18F]FDG PET/CT scans were performed in accordance with the EANM imaging guidelines [29] in a EANM research limited (EARL) PET/CT center [21], using version 2 of EARL (EARL2 [30]). Patients fasted for at least 6 h before intravenous [18F]FDG (3MBq/kg) injection. Their weight was checked on a calibrated scale [31]. All PET data were acquired in list mode about 1 h post-injection on a digital PET/CT (VEREOS, Philips Healthcare, 2017) with 90 s per bed position duration. Before each PET scan, a low-dose non-contrast-enhanced CT scan was acquired for attenuation correction and anatomical reference. An ultra-high definition (UHD) reconstruction was performed with 3D-ordered subset expectation maximization (OSEM), time-of-flight (TOF) and point spread function (PSF), using 4 iterations and 4 subsets, a 288 × 288 matrix and 2 × 2 × 2mm3 voxel size. Scatter and attenuation corrections were applied.

For each patient, two consecutive “standard PET” scans (PET1 and PET2) were performed during patient follow-up in clinical routine.

To obtain “AI PET” a second PET reconstruction was performed using half of acquisition data (45 instead of 90 s per bed position, with identical remaining reconstruction parameters) in baseline PET1 or in follow-up PET2 according to the study group as further specified.

This half-duration PET reconstruction was then treated by AI denoising using SubtlePET™ on a local in-house server (“AI PET”).

In study group 1, a supplemental AI PET was generated on PET2 whereas in group 2 a supplemental AI PET1 was created (Fig. 1).

Fig. 1
figure 1

Study design diagram of both study groups. “Standard” PET Response classifications between standard PET1 and standard PET2 were compared to “mixed” classifications between standard PET and AI PET (at follow-up AI PET 2: for study group 1; or at baseline AI PET 1 for group 2)

Full size image

Image analysis

Patients’ PET/CT images were randomized between 2 experienced nuclear medicine physicians and were reviewed in Syngo.via (version VB30A, Siemens Healthcare).

For each PET examination, SULmax and SULpeak of the most intense lesion (which could be different when considering the highest respective SULmax and SULpeak values) were recorded and taken into account for EORTC and PERCIST classifications, respectively. A semi-automatic segmentation method was performed using the 50% 3D-isocontour of the maximal pixel value on the UHD (± AI denoised) PET series. Both original UHD and EARL2 compliant Janma SUL values were used.

EARL2 compliant SUL values (SULmax−EARL2 and SULpeak−EARL2) were obtained by applying numerically to the SULUHD values a Gaussian post-filter with a full width half maximum of 3.6 mm (determined during the EARL2 accreditation process in standard PET) in Syngo.via (EQ.PET filter [32]). The same post-filter was used for AI PET as for standard PET.

For the EORTC response classification, the same lesion was considered on PET1 and PET2 for delta (Δ) SULmax calculation. For PERCIST, the lesion considered on PET1 and PET2 for ΔSULpeak computation could be different (the hottest on each scan, also named “target” lesion). Lesion size (short and long axis) of one target lesion with the highest SULpeak per examination was recorded.

Readers established “standard” response classifications between standard PET1 and standard PET2, considered the gold standard, and “mixed” response classifications between standard PET and AI PET (at baseline or follow-up).

For each patient, an interval of at least one month was respected between respectively standard and mixed classifications by the same reader (standard first in half of patients).

EORTC response classifications were established in all patients.

To be PERCISTUHD or EARL2 evaluable, SULpeak−UHD and EARL2 thresholds for the hottest target lesion in baseline PET1 (AI and/or standard) were respected, as well as other PERCIST criteria except for liver SULmean differences between PET1 and PET2 (concerning 7 patients with relative and/or absolute differences moderately exceeding 20% and/or 0.3SUL units between standard PET1 and 2) [9].

In each “standard” and “mixed” EORTC and PERCIST assessment, two separate classifications were established using either “UHD” SUL values (from original standard PET and AI PET) or “EARL2” [30] compliant SUL values besides visual evaluation of the UHD (standard and AI) PET images (EORTCUHD and EORTCEARL2 and, where applicable, PERCISTUHD and PERCISTEARL2 classifications).

EORTC and PERCIST response categories (applicable for treatments of more than one cycle) are defined as follows:

  • – Complete metabolic response (CMR): complete resolution of [18F]FDG uptake in all tumour volumes, indistinguishable from surrounding normal tissue for EORTC, and lower than liver SULmean and indistinguishable from background blood pool for PERCIST.

  • – Partial metabolic response (PMR): reduction of more than 25% in target tumour SULmax for EORTC, and of at least 30% and 0.8 SUL unit in target lesion SULpeak for PERCIST without increase (> 30%) in non-target lesion SULpeak and without new related tumour lesions. No significant increase (at least 20% in longest dimension) in tumour uptake extent (EORTC) or > 30% in size (PERCIST).

  • – Stable metabolic disease (SMD): change in tumour uptake of ≤ 25% for EORTC, or < 30% decrease and ≤ 30% increase for PERCIST without new lesions. No significant increase (at least 20% in longest dimension) in tumour uptake extent (EORTC) or > 30% in size (PERCIST).

  • – Progressive metabolic disease (PMD): greater than 25% increase in tumour [18F]FDG uptake (of a baseline target lesion) for EORTC and 30% and 0.8 SUL unit SULpeak increase for PERCIST, and/or appearance of new malignant lesions.

TLG was not properly taken into consideration for response classifications.

In clinical routine EORTCEARL2 PET response classifications had been reported and used for patient management.

In case of a discrepancy between any standard vs. mixed EORTC or PERCIST classification established for this study, an oncologist, specialized in the primary tumour, evaluated the potential impact on patient management the modified mixed vs. standard response classification could have had.

Statistical analysis

Quantitative variables were described by mean ± standard deviation and qualitative variables by frequencies and percentages. Concordance between “standard” PET therapy response classifications (i.e. between both standard PET1 and PET2) and “mixed” classifications (i.e. between AI and standard PET 1 or PET 2) was assessed by computing the unweighted Cohen’s kappa coefficient (κ) [33] with 95% confidence interval. EORTC and, where applicable, PERCIST classifications were both studied with respective assessable samples. Variations of semi-quantitative SUL measures between PET1 and PET2 (delta Δ = (SULPET2 -SULPET1) / SUL PET1) were compared between “standard” and “mixed” assessments by the non-parametric Wilcoxon Mann-Whitney test. Two-tailed testing with significance level of 5% (P < 0.05) was considered.

Analyses were conducted with R version 4.1.2.

Results

Patient inclusion and characteristics

One hundred and ten patients were included, 64 in group 1 (“standard” PET1 and “standard and AI” PET2) and 46 in group 2 (“standard and AI” PET1 and “standard” PET2). Their clinical and PET characteristics are described in Table 1.

Table 1 Patient characteristics
Full size table

The median interval between PET1 and PET2 was 126 days, interquartile range [96–183].

Image analysis

Overall, considering both study groups, 11/110 (10%) patients showed discordances in any of the mixed vs. standard EORTC and PERCISTEARL2 or UHD classifications, with a total of 28/397 (7%) discordant classifications.

The overall κ between mixed and standard classifications was 0.86 (95% confidence interval (CI) = 0.78–0.94), compatible with an almost perfect agreement.

In Fig. 2, we show in both study groups the distribution of standard and mixed EORTC and PERCISTEARL2 response classifications which were similar to classifications using SULUHD.

Fig. 2
figure 2

The distribution of “standard” - “mixed” EORTCEARL2 and PERCISTEARL2 response classifications. Group 1 (“standard” PET1 and “standard and AI” PET2) is represented in the upper part and Group 2 in the lower part (“standard and AI” PET1 and “standard” PET2). The absolute number and percentage of “standard” –“mixed” classification combinations are shown in the rectangles. CMR: complete metabolic response; PMR: partial metabolic response; PMD: progressive metabolic disease; SMD: stable metabolic disease. Concordant standard and mixed classifications are shown in white. Different discordant standard and mixed classifications are represented in grey scales

Full size image

In Table 2 we show lesion characteristics.

Table 2 The evolution of target lesion uptake and size between PET1 and PET2 in standard and mixed analysis
Full size table

Group 1: Standard PET1 and Standard and AI PET2

58/64 patients (91%) were also PERCIST evaluated (both standard and mixed PERCISTEARL2 and PERCISTUHD).

Overall, 57/64 (89%) patients were concordant in all performed mixed vs. standard response classifications (n = 51 being PERCIST evaluable), with 7 patients showing at least one discordance in any of the mixed vs. standard EORTC or PERCIST classifications (Fig. 2; Table 3).

Table 3 Discordant findings in response classifications between standard PET1 and standard PET2 (“standard”) vs. between standard PET1 and AI PET2 (“mixed”) in group 1
Full size table

Figure 3 shows a patient with concordant mixed vs. standard response classifications.

Fig. 3
figure 3

Representative PET images of a concordant patient included in group 1. MIP views of standard PET1 and standard and AI PET2 of a 58-year old woman referred for therapy response evaluation of metastatic breast cancer on fulvestrant and ribociclib. Classification was concordant ‘PMD’ in all standard and mixed PET response evaluations, due to multi-site nodal, bone and pulmonary progression (with new lesions) in PET2, along with a discordant response of other non-target lesions and a stable target breast lesion. PMD was confirmed at follow-up PET-CT 2 months later, after which treatment was switched into weekly paclitaxel

Full size image

For 6 out of 7 discordant patients in any classification, specialized oncologists established that these discordances would have had no impact on patients’ management. In one discordant patient classification change could have moderately impacted patient management. This patient previously diagnosed with stage IIIB T2bN2bM0 (TNM 9 [34,35,36] ) squamous cell lung carcinoma was re-evaluated by [18F]FDG PET/CT after 3 chemotherapy cycles. “Standard” PMR was misclassified as PMD in all “mixed” EORTC and PERCIST classifications due to a single indeterminate false positive liver focus in AI PET2. This would have led to a verification liver MRI and probable delay in initiation of radiotherapy.

Noteworthy, in these discordant patients, the same standard EORTCEARL2 classifications were obtained for this study as previously in clinical routine.

Group 2: Standard and AI PET1 and Standard PET2

30/46 patients (65%) were PERCISTEARL2 evaluated and 31/46 patients (67%) by PERCISTUHD. Overall, up to 15 patients were PERCIST (EARL2 and/or UHD) excluded in the mixed PET group (due to insufficient baseline AI PET1 uptake) vs. up to 12 patients in the gold standard group (due to insufficient standard PET1 uptake).

42/46 (91%) patients were concordant in all performed mixed vs. standard EORTC and PERCIST classifications, of whom 29 being PERCIST evaluable. Table 4 describes the discordant findings.

Table 4 Discordant findings in response classifications between standard PET1 and standard PET2 (“standard”) vs. between AI PET1 and standard PET2 (“mixed”) in group 2
Full size table

One mixed vs. standard EORTCEARL2 classification change could have impacted patient management (Fig. 4).

Fig. 4
figure 4

Representative PET images of a discordant patient included in group 2. MIP views and an axial fused PET/CT image of a 66 year-old woman before (PET1) and after (PET2) 4 cycles of chemotherapy (oxaliplatin) for presacral recurrence (thin red arrows) of recto-sigmoidal cancer. Only standard EORTCEARL2 classification was ‘PMD’ which had been reported previously in clinical routine. All the other classifications were ‘SMD’. ‘PMD’ was based on a slight increase of target lesion SULmax−EARL2 (+ 25.4%) in standard PET2 vs. standard PET1 versus only + 18.9% in standard PET2 vs. AI PET1, without any significant changes in size, MATV or TLG; nor new or other clearly suspicious progressive lesions. ‘PMD’ induced a chemotherapy switch (into fluorouracil with irinotecan) in clinical routine, which would have been omitted in case of ‘SMD’

Full size image

Standard EORTCEARL2 classifications of discordant patients were identical to classifications previously reported in clinical routine.

Discussion

In this study, we evaluated whether introducing AI-based [18F]FDG PET denoising and comparing mixed standard and AI denoised half-duration PET (at baseline or follow-up) has an impact on EORTC and PERCIST solid tumour therapy response classifications.

The standard classifications between two standard PET, as a gold standard, were challenged against mixed classifications.

Indeed, this question is relevant during a transition period or long-lasting if patients keep on undergoing randomly assigned AI denoised or standard PET during follow-up (use of AI in a part of the routine PET population which is a common scenario related to economical and organisational factors).

Overall, mixed vs. standard response classification discordances in any of the 4 classifications (EORTC or PERCISTEARL2 or UHD) were found in a minority of this study population, i.e. in 11/110 (10%) patients, and in 7% of all classifications.

This discordance percentage seems less pronounced than in previous response evaluation studies with inhomogeneous PET reconstructions, i.e. between mixed OSEM and OSEM + PSF ± TOF PET vs. between 2 OSEM PET in analog PET/CT [19, 20]. Also in contrast with these studies and despite previously found higher differences in SULmaxvs. in SULpeak between standard and AI PET [24], classification discordances in this study population were not less pronounced but comparable in PERCIST vs. in EORTC classifications (in each 7 discordant patients, when PERCIST evaluable). Moreover the number of discrepancies was not less pronounced when using EARL2 compliant vs. UHD SUL values, as 2 more patients became discordant in EARL2 classifications (n = 10 vs. 8).

Except for one discordant patient caused by an indeterminate false positive focus in AI PET2, discordances were mainly induced by lesion SUL and tumour-to-background uptake ratio changes.

It is noteworthy that 89/110 (82%) patients were included in the PERCIST analysis (n = 88 in both PERCISTEARL2 and in PERCISTUHD, predominantly in the first study group: n = 58).

Five out of 11 (45%) of discordant patients were discordant in all 4 classification systems while 4/11 (36%) discordant patients (one excluded for PERCIST) had a difference only in one classification system (EORTC or PERCIST, EARL2 or UHD).

More important, classification discordance could have had an impact on patient management in up to 2/110 (2%) patients.

Study drawbacks.

No reference standard for evaluating the accuracy of the PET response classification was available or reliable in most cases (like histopathology or a combination of other radiological and clinico-biological data), as is nevertheless common in clinical routine. Also patient management and treatment modifications were based in the majority of cases mainly on standard PET 1 and 2 comparison as reported in clinical routine. Noteworthy, breast cancer was overrepresented in this study population.

SUL changes are a continuum and response categorisation based on SUL cut-offs is an inherent limitation of response classifications. SUL percentage changes have also been analysed in this study, and were not always significantly different between standard and mixed scenario’s (in group 1 only for Δ SULmax−UHD).

All patients were fully EORTC evaluated, although a minority of patients showed only moderate-uptake (often small) lesions. This is however a common routine practice scenario and no strict SULmax threshold for EORTC is defined [7].

The number of in particular PERCIST evaluated patients was limited in the second study group.

Moreover, the liver SULmean criteria for PERCIST evaluability were not taken into account for the sake of simplification and being also less applicable when comparing an AI PET to a standard PET. As demonstrated in previous studies, AI was found to slightly increase liver SULmean [24, 28, 37].

We did not study the impact on treatment response evaluation when comparing two AI denoised PET vs. two standard PET and vs. two mixed PET, as few patients were includable for the scenario with both AI PET1 and AI PET2. However, as demonstrated in a previous study, intra-class correlation coefficients for lesion SULmax and SULpeak between standard PET and AI PET respectively were very high (> 0.97) [24], so little difference is expected in response classifications between 2 AI vs. between 2 standard PET.

We did not properly take into account changes in metabolically active tumour volume (MATV) and TLG. In PERCIST, uncertainty exists about how to measure them [8, 9]. Accurate (automatic) whole tumour burden segmentation could be more meaningful than of only target lesions, and the added classification value of these measures needs to be further studied.

AI can also be used for whole tumour burden segmentation [38], with several commercial viewing servers disposing of this functionality.

To the best of our knowledge, this is the first study on the impact on PET therapy response classification of introducing AI PET denoising and of the mixed use of AI-denoised half-duration and non-AI [18F]FDG PET during follow-up, which is a common scenario in routine practice.

It sets a basis for further investigation and understanding of the clinical implications of introducing AI denoising in PET.

Conclusions

In 10% of patients any “mixed” [18F]FDG PET EORTC or PERCIST response classification (between a standard PET and an AI denoised half-duration PET) was discordant with the corresponding “standard” classification (between two standard PET).

However, in only 2% of patients this discordance could have had an impact on patient management.

Therefore, although comparison of standardised PET images (with identical imaging protocols) is always preferable, the mixed comparison of an AI denoised half-duration PET and non AI full-duration PET in the therapeutic assessment process seems clinically tolerable.

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Abbreviations

[18F]FDG:

[18F]fluorodeoxyglucose

PET:

Positron emission tomography

AI:

Artificial intelligence

EORTC:

European Organisation for Research and Treatment of Cancer (PET response classification)

PERCIST:

PET evaluation response criteria in solid tumours

UHD:

Ultra-high definition

EARL2:

EANM research limited version 2

SUV:

Standardized uptake values

SUL:

Lean body mass corrected standardized uptake values

SULmax :

The highest (maximal) lean body mass corrected standardized uptake pixel value

SULpeak :

Mean SUL in a spherical 1-cm3 (1.2 cm diameter) volume of interest (VOI) around the hottest focus i.e. maximal pixel value

TLG:

Total lesion glycolysis

MATV:

Metabolically active tumour volume

OSEM:

Ordered subset expectation maximization

TOF:

Time-of-flight

PSF:

Point spread function

CMR:

Complete metabolic response

PMR:

Partial metabolic response

SMD:

Stable metabolic disease

PMD:

Progressive metabolic disease

CI:

Confidence interval

BMI:

Body mass index

RT:

Radiation therapy

References

  1. Global Burden of Disease (GBD). [Internet]. [cited 2023 Aug 3]. https://www.healthdata.org/research-analysis/gbd.

  2. Cancer Progress Timeline [Internet]. ASCO. [cited 2023 Aug 3]. https://old-prod.asco.org/news-initiatives/cancer-progress-timeline.

  3. Ben-Haim S, Ell P. 18F-FDG PET and PET/CT in the evaluation of Cancer Treatment Response. J Nucl Med. 2009;50:88–99. https://doi.org/10.2967/jnumed.108.054205.

    Article  PubMed  Google Scholar 

  4. Gallamini A, Zwarthoed C, Borra A. Positron Emission Tomography (PET) in Oncology. Cancers. 2014;6:1821–89. https://doi.org/10.3390/cancers6041821.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Hirata K, Tamaki N. Quantitative FDG PET Assessment for Oncology Therapy. Cancers. 2021;13:869. https://doi.org/10.3390/cancers13040869.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  6. Aide N, Hicks RJ, Le Tourneau C, Lheureux S, Fanti S, Lopci E. FDG PET/CT for assessing tumour response to immunotherapy. Eur J Nucl Med Mol Imaging. 2019;46:238–50. https://doi.org/10.1007/s00259-018-4171-4.

    Article  PubMed  Google Scholar 

  7. Young H, Baum R, Cremerius U, Herholz K, Hoekstra O, Lammertsma AA, et al. Measurement of clinical and subclinical tumour response using [18F]-fluorodeoxyglucose and positron emission tomography: review and 1999 EORTC recommendations. Eur J Cancer. 1999;35:1773–82. https://doi.org/10.1016/S0959-8049(99)00229-4.

    Article  CAS  PubMed  Google Scholar 

  8. Joo Hyun O, Lodge MA, Wahl RL. Practical PERCIST: a simplified guide to PET response criteria in solid tumors 1.0. Radiology. 2016;280:576. https://doi.org/10.1148/radiol.2016142043.

    Article  Google Scholar 

  9. Wahl RL, Jacene H, Kasamon Y, Lodge MA. From RECIST to PERCIST: evolving considerations for PET response criteria in solid tumors. J Nucl Med. 2009;50:S122–50. https://doi.org/10.2967/jnumed.108.05730.

    Article  Google Scholar 

  10. Pinker K, Riedl C, Weber WA. Evaluating tumor response with FDG PET: updates on PERCIST, comparison with EORTC criteria and clues to future developments. Eur J Nucl Med Mol Imaging. 2017;44:55–66. https://doi.org/10.1007/s00259-017-3687-3.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Weber WA. Use of PET for Monitoring Cancer Therapy and for Predicting Outcome. J Nucl Med. 2005;46:983–95.

    CAS  PubMed  Google Scholar 

  12. Weber WA, Figlin R. Monitoring Cancer Treatment with PET/CT: does it make a difference? J Nucl Med. 2007;48:S36–44.

    Google Scholar 

  13. Kelloff GJ, Hoffman JM, Johnson B, Scher HI, Siegel BA, Cheng EY, et al. Progress and Promise of FDG-PET Imaging for Cancer Patient Management and Oncologic Drug Development. Clin Cancer Res. 2005;11:2785–808. https://doi.org/10.1158/1078-0432.CCR-04-2626.

    Article  CAS  PubMed  Google Scholar 

  14. Juweid Malik E, Cheson Bruce D. Positron-Emission Tomography and Assessment of Cancer Therapy. N Engl J Med. 2006;354:496–507. https://doi.org/10.1056/NEJMra050276.

    Article  CAS  PubMed  Google Scholar 

  15. Dimitriou F, Lo SN, Tan AC, Emmett L, Kapoor R, Carlino MS, et al. FDG-PET to predict long-term outcome from anti-PD-1 therapy in metastatic melanoma. Ann Oncol. 2022;33:99–106. https://doi.org/10.1016/j.annonc.2021.10.003.

    Article  CAS  PubMed  Google Scholar 

  16. Boellaard R. Standards for PET image acquisition and quantitative data analysis. J Nucl Med. 2009;50. https://doi.org/10.2967/jnumed.108.057182. :11S-20S.

  17. Kinahan PE, Perlman ES, Sunderland JJ, Subramaniam R, Wollenweber SD, Turkington TG, et al. The QIBA Profile for FDG PET/CT as an imaging biomarker measuring response to Cancer Therapy. Radiology. 2020;294:647–57. https://doi.org/10.1148/radiol.2019191882.

    Article  PubMed  Google Scholar 

  18. Lodge MA. Repeatability of SUV in oncologic 18F-FDG PET. J Nucl Med. 2017;58:523–32. https://doi.org/10.2967/jnumed.116.186353.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  19. Quak E, Roux P-YL, Lasnon C, Robin P, Hofman MS, Bourhis D, et al. Does PET SUV Harmonization Affect PERCIST Response classification? J Nucl Med. 2016;57:1699–706. https://doi.org/10.2967/jnumed.115.17198.

    Article  CAS  PubMed  Google Scholar 

  20. Lasnon C, Quak E, Le Roux P-Y, Robin P, Hofman MS, Bourhis D, et al. EORTC PET response criteria are more influenced by reconstruction inconsistencies than PERCIST but both benefit from the EARL harmonization program. EJNMMI Phys. 2017;4:17. https://doi.org/10.1186/s40658-017-0185-4.

    Article  PubMed  PubMed Central  Google Scholar 

  21. Aide N, Lasnon C, Veit-Haibach P, Sera T, Sattler B, Boellaard R. EANM/EARL harmonization strategies in PET quantification: from daily practice to multicentre oncological studies. Eur J Nucl Med Mol Imaging. 2017;44:17–31. https://doi.org/10.1007/s00259-017-3740-2.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Arabi H, AkhavanAllaf A, Sanaat A, Shiri I, Zaidi H. The promise of artificial intelligence and deep learning in PET and SPECT imaging. Physica Med. 2021;83:122–37. https://doi.org/10.1016/j.ejmp.2021.03.008.

    Article  Google Scholar 

  23. SubtlePET | Subtle Medical [Internet]. [cited 2021 Nov 17]. https://subtlemedical.com/subtlepet/.

  24. Weyts K, Lasnon C, Ciappuccini R, Lequesne J, Corroyer-Dulmont A, Quak E et al. Artificial intelligence-based PET denoising could allow a two-fold reduction in [18F]FDG PET acquisition time in digital PET/CT. Eur J Nucl Med Mol Imaging [Internet]. 2022 [cited 2022 Jun 10]; https://doi.org/10.1007/s00259-022-05800-1.

  25. Quak E, Weyts K, Jaudet C, Prigent A, Foucras G, Lasnon C. Artificial intelligence-based 68Ga-DOTATOC PET denoising for optimizing 68Ge/68Ga generator use throughout its lifetime. Frontiers. 2023;10:1137514. https://doi.org/10.3389/fmed.2023.1137514.

    Article  Google Scholar 

  26. Katsari K, Penna D, Arena V, Polverari G, Ianniello A, Italiano D, et al. Artificial intelligence for reduced dose 18F-FDG PET examinations: a real-world deployment through a standardized framework and business case assessment. EJNMMI Phys. 2021;8:25. https://doi.org/10.1186/s40658-021-00374-7.

    Article  PubMed  PubMed Central  Google Scholar 

  27. Jaudet C, Weyts K, Lechervy A, Batalla A, Bardet S, Corroyer-Dulmont A. The Impact of Artificial Intelligence CNN based denoising on FDG PET Radiomics. Front Oncol. 2021;3136. https://doi.org/10.3389/fonc.2021.692973.

  28. Weyts K, Quak E, Licaj I, Ciappuccini R, Lasnon C, Corroyer-Dulmont A, et al. Deep Learning Denoising improves and homogenizes patient [18F]FDG PET Image Quality in Digital PET/CT. Diagnostics. 2023;13:1626. https://doi.org/10.3390/diagnostics13091626.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  29. Boellaard R, Delgado-Bolton R, Oyen WJG, 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;42:328–54. https://doi.org/10.1007/s00259-014-2961-x.

    Article  CAS  PubMed  Google Scholar 

  30. Kaalep A, Sera T, Rijnsdorp S, Yaqub M, Talsma A, Lodge MA, et al. Feasibility of state of the art PET/CT systems performance harmonisation. Eur J Nucl Med Mol Imaging. 2018;45:1344–61. https://doi.org/10.1007/s00259-018-3977-4.

    Article  PubMed  PubMed Central  Google Scholar 

  31. Lasnon C, Houdu B, Kammerer E, Salomon T, Devreese J, Lebasnier A, et al. Patient’s weight: a neglected cause of variability in SUV measurements? A survey from an EARL accredited PET centre in 513 patients. Eur J Nucl Med Mol Imaging. 2016;43:197–9. https://doi.org/10.1007/s00259-015-3214-3.

    Article  PubMed  Google Scholar 

  32. Quak E, Le Roux P-Y, Hofman MS, Robin P, Bourhis D, Callahan J, et al. Harmonizing FDG PET quantification while maintaining optimal lesion detection: prospective multicentre validation in 517 oncology patients. Eur J Nucl Med Mol Imaging. 2015;42:2072–82. https://doi.org/10.1007/s00259-015-3128-0.

    Article  PubMed  PubMed Central  Google Scholar 

  33. Landis JR, Koch GG. An application of hierarchical Kappa-type statistics in the Assessment of Majority Agreement among multiple observers. Biometrics. 1977;33:363–74. https://doi.org/10.2307/2529786.

    Article  CAS  PubMed  Google Scholar 

  34. Van Schil PE, Asamura H, Nishimura KK, Rami-Porta R, Kim YT, Bertoglio P et al. The International Association for the Study of Lung Cancer Lung Cancer Staging Project: Proposals for the Revisions of the T-Descriptors in the Forthcoming Ninth Edition of the TNM Classification for Lung Cancer. Journal of Thoracic Oncology [Internet]. 2023 [cited 2024 Mar 18]; https://www.sciencedirect.com/science/article/pii/S1556086423024115. https://doi.org/10.1016/j.jtho.2023.12.006.

  35. Huang J, Osarogiagbon RU, Giroux DJ, Nishimura KK, Bille A, Cardillo G et al. The International Association for the Study of Lung Cancer Staging Project for Lung Cancer: Proposals for the Revision of the N Descriptors in the Forthcoming Ninth Edition of the TNM Classification for Lung Cancer. Journal of Thoracic Oncology [Internet]. 2023 [cited 2024 Mar 18]; https://www.sciencedirect.com/science/article/pii/S1556086423023109. https://doi.org/10.1016/j.jtho.2023.10.012.

  36. Fong KM, Rosenthal A, Giroux DJ, Nishimura KK, Erasmus J, Lievens Y et al. The IASLC Lung Cancer Staging Project: Proposals for the Revision of the M Descriptors in the Forthcoming 9th edition of the TNM Classification of Lung Cancer. Journal of Thoracic Oncology [Internet]. 2024 [cited 2024 Mar 18]; https://www.sciencedirect.com/science/article/pii/S1556086424000601. https://doi.org/10.1016/j.jtho.2024.01.019.

  37. Bonardel G, Dupont A, Decazes P, Queneau M, Modzelewski R, Coulot J, et al. Clinical and phantom validation of a deep learning based denoising algorithm for F-18-FDG PET images from lower detection counting in comparison with the standard acquisition. EJNMMI Phys. 2022;9:36. https://doi.org/10.1186/s40658-022-00465-z.

    Article  PubMed  PubMed Central  Google Scholar 

  38. Blanc-Durand P, Jégou S, Kanoun S, Berriolo-Riedinger A, Bodet-Milin C, Kraeber-Bodéré 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;48:1362–70. https://doi.org/10.1007/s00259-020-05080-7.

    Article  CAS  PubMed  Google Scholar 

Download references

Acknowledgements

not applicable.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Author information

Authors and Affiliations

  1. Nuclear Medicine Department, François Baclesse Comprehensive Cancer Centre, UNICANCER, Caen, 3 Avenue du General Harris, BP 45026, Caen Cedex 5, 14076, France

    Kathleen Weyts & Charline Lasnon

  2. Biostatistics Department, François Baclesse Comprehensive Cancer Centre, UNICANCER, Caen, France

    Justine Lequesne

  3. Medical Oncology Department, François Baclesse Comprehensive Cancer Centre, UNICANCER, Caen, France

    Alison Johnson, Hubert Curcio, Aurélie Parzy & Elodie Coquan

  4. UNICAEN, INSERM 1086 ANTICIPE, Normandy University, Caen, France

    Charline Lasnon

Authors
  1. Kathleen Weyts
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Justine Lequesne
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Alison Johnson
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Hubert Curcio
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Aurélie Parzy
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Elodie Coquan
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Charline Lasnon
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

JL, KW and CL contributed to methodology and formal analysis. KW, CL, AJ, HC, AP and EQ provided resources. Conceptualization, validation and original manuscript drafting were performed by KW. All authors reviewed, commented on and approved the work and final manuscript.

Corresponding author

Correspondence to Kathleen Weyts.

Ethics declarations

Ethics approval

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee at the François Baclesse Comprehensive Cancer Centre and registered with the French Health Data Hub under reference F20220712121936.

Consent for publication

The authors affirm that human research participants provided informed consent for publication of the images in Figs. 3 and 4.

Consent to participate

Informed consent was obtained from all individual participants included in the study.

Competing interests

The authors have no relevant financial or non-financial interests to disclose.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Weyts, K., Lequesne, J., Johnson, A. et al. The impact of introducing deep learning based [18F]FDG PET denoising on EORTC and PERCIST therapeutic response assessments in digital PET/CT. EJNMMI Res 14, 72 (2024). https://doi.org/10.1186/s13550-024-01128-z

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13550-024-01128-z

Keywords