Evaluation of an artificial intelligence method for lesion segmentation of baseline FDG PET studies of DLBCL patients

Evaluation of an artificial intelligence method for lesion segmentation of baseline FDG PET studies of DLBCL patients | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Evaluation of an artificial intelligence method for lesion segmentation of baseline FDG PET studies of DLBCL patients Maria C. Ferrández, Sandeep S. V. Golla, Sara C. A. De Visser, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6294601/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background. The aim of this study is to investigate the use of an artificial intelligence (AI) method, LIONZ, in combination with an intensity-based threshold method, SUV4.0, for the automatic selection and segmentation of diffuse large B cell lymphoma (DLBCL) lymphoma lesions. Methods. 296 DLBCL 18 F-FDG PET scans were analyzed. Metabolic tumor volume, peak standardized uptake value (SUVpeak) and, maximum distance from the bulkiest lesion to another lesion (Dmaxbulk) were extracted from the LIONZ and LIONZ SUV4 segmentations and compared to those extracted from SUV4.0 segmentations using Pearson correlation (p < 0.05) and Bland-Altman plots. Segmentation performance was assessed using the Dice similarity coefficient (DSC) with SUV4.0 segmentation as a reference. A prediction model which includes MTV, SUVpeak, Dmaxbulk, age and performance status was used to predict the probability of 2 year time to progression using the parameters extracted from the LIONZ, LIONZ SUV4 and SUV4.0 segmentations. Association of probabilities was evaluated using Pearson correlation (p < 0.05) and Bland-Altman. The area under (AUC) the curve was used to assess and compare the performance of both methods. Results. The median DSC (interquartile range) for LIONZ when compared to SUV4.0 was of 0.77 (0.64–0.84) and for LIONZ SUV4 of 0.87 (0.80–0.93). MTV, SUVpeak and Dmaxbulk from both the LIONZ and LIONZ SUV4 were highly correlated to the SUV4.0 segmentations derived parameters (R ≥ 0.80, p < 0.0001). LIONZ SUV4 reduced overestimation of segmented areas and LIONZ SUV4 MTV showed a stronger agreement with that of SUV4.0 compared to LIONZ (0.99 and 0.80 respectively, p < 0.0001). The prediction model yielded an AUC of 0.74, 0.78 and 0.79 when using segmentations from LIONZ, LIONZ SUV4 and SUV4.0 respectively. The predicted probabilities yielded by the models using the LIONZ and LIONZ SUV4 segmentations were also highly correlated with those of SUV4.0 segmentation (0.9 and 0.96 respectively, p < 0.0001). Conclusion. LIONZ SUV4 segmentations highly overlapped with those of SUV4.0. LIONZ SUV4 led to a stronger agreement of PET parameters and predictions with SUV4.0 compared to LIONZ. Overall, LIONZ SUV4 is a suitable method for DLBCL lesion segmentation and potentially decreases reader-variability compared to threshold only based segmentation methods. Nuclear Medicine & Medical Imaging Artificial Intelligence and Machine Learning Hematology artificial intelligence tumor segmentation diffuse large b cell lymphoma PET imaging Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 INTRODUCTION Diffuse large B cell lymphoma (DLBCL) is the most common type of non-Hodgkin lymphoma, accounting for 30% of all non-Hodgkin lymphoma diagnoses (1). It is characterized by diffused patterns of growth and for being an extremely heterogeneous disease, which hinders treatment planning. The International Prognostic Index (IPI), and variations of these parameters, have served clinicians in risk stratification and guided treatment strategies for DLBCL patients (2–4). However, advancements in diagnostics and therapeutic treatments have significantly improved DLBCL outcomes particularly for high-risk patients (1). As a result, IPI’s effectiveness in reliably predicting treatment failure has declined (4). 18 F-Fluorodeoxyglucose ( 18 F-FDG) positron emission tomography (PET) in combination with computed tomography (CT) is a standard tool for the prognostic evaluation of DLBCL. There is evidence that parameters quantified through 18 F-FDG PET/CT imaging have predictive potential in lymphoma, such as the metabolic tumor volume (MTV) and the maximum distance between the bulkiest lesion and any other lesion (Dmaxbulk) (5). Segmentation of tumors is required to quantify such parameters from the PET images. Standard uptake value (SUV) thresholding methods are most commonly used for tumor segmentation. In the case of DLBCL, SUV4.0 is the segmentation method which has shown to be most robust (6). However, these semi-automatic methods require manual input and/or lesion selection thereby becoming time-consuming and sensitive to reader variability. In this study we used LIONZ (7), an externally developed artificial intelligence (AI) tool, to automatically detect DLBCL lesions. We further optimized LIONZ by combining it with SUV4.0 thresholding (LIONZ SUV4 ). The segmentations derived from LIONZ and LIONZ SUV4 were compared to those of SUV4.0. SUV4.0 has been referred to as the preferred segmentation threshold in DLBCL, as shown by Barrington et al . (6) and by the international lymphoma MTV benchmark (8). The benchmark established that small lesions with a volume lower than 3 mL needs to be manually included in the SUV4.0 segmentation. Additionally, MTV, SUVpeak and Dmaxbulk values were extracted and analysed for all the three segmentation methods. Furthermore, we assessed the predictive power of PET parameters derived from the LIONZ and LIONZ SUV4 segmentations using the prediction model developed by Eertink et al. (5). The model is a logistic regression outcome prediction model which utilizes MTV, Dmaxbulk and SUVpeak parameters extracted from SUV4.0 segmentations together with two clinical parameters, age and World Health Organization (WHO) performance status. The model outcome is a dichotomous value equivalent to the probability of time to progression (TTP) within or longer than 2 years in DLBCL patients. METHODS Data A total of 317 baseline DLBCL 18 F-FDG PET/CT scans from the HOVON-84 trial were available for this study (EudraCT, 2006-005, 174 − 42) (9). This study was approved by institutional review boards and all included patients provided informed consent. The use of all data within the HOVON-84 database has been approved by the institutional review board of the VU University Medical Center (JR /20140414). From the 317 patients, 7 were lost to follow-up within 2 years and 14 other patients died within 2 years of unrelated reasons. This led to a total of 296 DLBCL patients included in this study. Quality control of scans was conducted following the criteria set by the European Association of Nuclear Medicine (10) and is fully described elsewhere (5). The scans included in this study are compliant with EARL1 standards. SUV4.0 segmentation Lesions were delineated through an automated preselection defined by a SUV ≥ 4.0 and a volume threshold of ≥ 3 mL. Non-tumor regions were removed, and lesions smaller than 3 mL were interactively added with one-click per lesion. If tumor regions were adjacent to high intensity non-tumor regions (e.g., kidney, bladder), these were manually corrected. Further details on the delineation methods and workflow are provided in the international lymphoma MTV benchmark for segmentation (8). All scans were reviewed by a nuclear medicine physician (GJCZ), and delineations were performed under his supervision. AI lesion detection and segmentation LIONZ is a deep learning tool developed by the ENHANCE PET ( https://enhance.pet/ ) group at Medical University of Vienna, Austria. LIONZ was trained on over 1000 18 F-FDG PET/CT scans of patients with multiple cancer types (lymphoma, melanoma and lung cancer) and healthy subjects. LIONZ is open-source and is available for download at Zenodo (7). Besides analyzing the segmentation performance of LIONZ, we developed LIONZ SUV4 which combines LIONZ and SUV4.0 into the segmentation workflow for the same purpose. This new method uses LIONZ as a lesion detection and selection step and is followed by SUV4.0 threshold region growing. In a second step, one-click corrections are used to manually remove injection spots, spots outside of the body and any segmented high uptake organ adjacent to tumor regions (brain, bladder and/or kidney). These corrections are limited to removal of regions and no missed lesions were manually added unlike for the SUV4.0 method that require reader selection of lesions smaller than 3 mL (8). SUV4.0, LIONZ and LIONZ SUV4 methods were implemented or embedded using the same software tool ACCURATE (11). ACCURATE enables the quantitative analysis of the PET scans and computation of PET parameters. We used ACCURATE to calculate MTV, SUVpeak and Dmaxbulk for each PET scan. Prediction model The prediction model was first developed using the HOVON-84 trial by Eertink et al. (5). The following features are included in this model: MTV, SUVpeak, Dmaxbulk, age and WHO performance status. These features were used as predictors in a logistic regression model to predict the probability of 2-year TTP for each patient. For the definition of TTP, patients who died within 2 years starting at the time of baseline scan without signs of progression were excluded from the analysis. Statistical analysis The dice similarity coefficient (DSC) was used to compare the performance of the three different segmentations. Pearson correlation and Bland-Altman plots were used to assess the relation between segmentation methods for all three PET parameters: MTV, SUVpeak and Dmaxbulk values. In order to assess the predictive potential the PET parameters derived from the LIONZ and LIONZ SUV4 segmentations, we replaced the features’ values in the logistic regression of the Clinical PET model by those of the LIONZ and LIONZ SUV4 based MTV, SUVpeak and Dmaxbulk values. We used the area under the curve (AUC) to assess the model’s performance. The Delong test was used to compare AUC values between methods (12). In addition, a risk classification analysis of the predicted TTP values yielded by the prediction model using the three different segmentations was also performed. Using the predicted TTP values derived from the model using SUV4.0 segmentations, 20% of patients with the highest predicted values were classified as high risk, 50% of the patients with the lowest predicted values in the low risk group and the rest as the intermediate risk group. The risk classifications from the prediction model using SUV4.0 segmentations was used as reference and, the effect on the risk classifications when using the predicted TTP values from other segmentations (LIONZ and LIONZ SUV4 ) was evaluated. RESULTS Early in the analysis, we found that LIONZ consistently identified areas beyond the actual lesion borders and labeling surrounding healthy tissue areas as part of the lesion. This led to an overestimation of the tumor size and volume (i.e. over-segmentation). Moreover, in cases with largely disseminated tumors and smooth tumor borders, LIONZ tended to under-segment lesions leaving out regions of interest that should be included in the segmentation (i.e. under-segmentation). Therefore, combining LIONZ with a SUV4.0 threshold (LIONZ SUV4 ) was evaluated as an option to correct the under/over-segmentation by LIONZ. In the LIONZ SUV4 method, LIONZ is used as a lesion detection step, followed by SUV4.0 threshold based region shrinking/growing, which compensates for any over/under-segmentation from LIONZ. Some examples that illustrate how LIONZ SUV4 led to a more accurate segmentation compared to LIONZ are given in Figures 1 and 2. Occasionally selected high uptake organs such as brain, bladder and/or kidneys adjacent to tumor regions were also manually removed for both LIONZ and LIONZ SUV4 . The most complicated cases involved removing the kidneys slice by slice from the PET images but these accounted for only 7 cases in total (2%). Examples of these corrections are given in Figure 4. Unlike the corrections for the SUV4.0 segmentation, no lesions were manually added. The DSC was calculated for all segmentations from LIONZ and LIONZ SUV4 with SUV4.0 as the reference segmentation. The median DSC and interquartile range (IQR) resulted in 0.77 (0.64 - 0.84) for LIONZ and 0.87 (0.80 - 0.93) for LIONZ SUV4 . There were 6 segmentations for which the DSC was equal to 0. These 6 cases corresponded to very small lesions (<3ml) in the SUV4.0 segmentations. For 4 of these cases, LIONZ failed in detecting any lesions. Unlike for LIONZ and LIONZ SUV4 , when using SUV4.0 segmentation, small lesions (<3ml) are manually added as indicated in (8). MTV values were highly correlated for LIONZ and SUV4.0 (0.80, p<0.0001). Generally, LIONZ overestimates tumor volume compared to SUV4.0, as we see in Figure 4 and Supplemental Figure 1A. The cases were LIONZ underestimates MTV correspond to those patients with largely diffused and non-defined tumor regions. When using LIONZ SUV4 we found that MTV values were more aligned with those of SUV4.0, with a higher correlation (0.98, p<0.0001). Moreover, the Bland-Altman plot in Supplemental Figure 1B shows an improved agreement of the MTV values when using LIONZ SUV4 compared to LIONZ. Regarding SUVpeak, the correlation was similar to that of MTV (0.98, p-value <0.0001). This was the case for both LIONZ and LIONZ SUV4 . LIONZ and LIONZ SUV4 SUVpeak values which led to 0 correspond to patients in which tumors were not detected at all (Figure 5). A slight overestimation of SUVpeak values is illustrated in the Bland-Altman plots for both LIONZ and LIONZ SUV4 (Supplemental Figure 2). Pearson correlation for Dmaxbulk was of 0.83 (p-value < 0.0001) for LIONZ and SUV4.0 values. For multiple patients, LIONZ identified small lesions that do not appear in the SUV4.0 segmentations which leads to large differences in Dmaxbulk (Figure 6). This seems to be improved when using LIONZ SUV4 (Supplemental Figure 3). The correlation was slightly higher when comparing LIONZ SUV4 and SUV4.0 (0.89, p<0.0001). Cases with the largest differences were usually patients with multiple small lesions largely spread. A summary of statistics for these parameters is given in Table 1. TTP probabilities were calculated with the prediction model using both the LIONZ and LIONZ SUV4 segmentations and compared to those of the original model (i.e. using SUV4.0 segmentations). The Pearson correlation was of 0.90 and 0.96 respectively (p-value < 0.0001, Figure 7A). The agreement between probabilities increased when using LIONZ SUV4 compared to LIONZ segmentations as shown in the Bland-Altman plot in Supplemental Figure 4. The AUC for the prediction model derived from the SUV4.0 segmentations is 0.79. Replacing the SUV4.0 parameters with those derived from the LIONZ segmentations yielded an AUC of 0.74 and for LIONZ SUV4 , the model achieved an AUC of 0.78 (Figure 7B). The Delong test found the AUC values for the model derived from LIONZ segmentations and the model derived from SUV4.0 segmentations to be statistically different (P 0.05). In Tables 2 and 3 we illustrate the changes in the risk classification from the model using SUV4.0 segmentations to the models using LIONZ and LIONZ SUV4 segmentations. Regarding the model that used LIONZ segmentations, the number of patients that remained in the same risk groups as the model derived from the SUV4.0 segmentation was 48 out of 59, 72 out of 90 and 132 out of 147 for the high, intermediate, and low risk group respectively. In the case of LIONZ SUV4 segmentation, 44 out of 59, 74 out of 90, and 145 out of 147 remained in the high, intermediate and low risk groups respectively. DISCUSSION AI based models are quickly advancing in the PET imaging field with multiple applications. Automatic segmentation is one of the most promising tasks (13). However, segmentation of lymphoma can be challenging due to the heterogeneity of the disease. In this study we investigated the potential of LIONZ, an AI method for the detection of DLBCL lesions. Moreover, we developed LIONZ SUV4 , which combined LIONZ and SUV4.0 in an effort to overcome LIONZ tendency to over/under-segment DLBCL lesions. We found that the segmentations yielded by LIONZ and LIONZ SUV4 exhibited a high level of similarity with those of SUV4.0 with a median DSC of 0.77 and 0.87, respectively. Moreover, the PET extracted parameters derived from LIONZ and LIONZ SUV4 segmentations are highly correlated with those of SUV4.0. These findings suggest that LIONZ itself is capable of identifying suspicious uptake sites and accurately discarding physiological regions with high uptake for nearly all patients. The modified version including SUV4.0 (LIONZ SUV4 ) is meant to compensate for the over-/under segmentation tendency of LIONZ, as shown in Figs. 1 and 2 . Overall, LIONZ SUV4 shows a stronger agreement of PET parameters with SUV4.0. Manual corrections were added to the pipeline to remove non-tumor regions. These corrections are mostly one-click interactions. In general, LIONZ accurately identified smaller lesions (< 3 ml) in the PET images, therefore we did not include any corrections involving manually adding suspected missed lesions, unlike for SUV4.0 corrections where many lesions would have been missed and therefore readers are required to manually add missed lesions as indicated in (8). The manual addition of missed lesions inevitably increases the number of reader interactions which is one of the reasons why SUV4.0 segmentations suffer from inter-reader variability. The new approach proposed by LIONZ SUV4 , requires fewer interactions by decreasing the discrepancies regarding non-detected smaller lesions (< 3ml), eventually decreasing inter-reader variability. These findings suggest that LIONZ is an efficient tool for the identification of lesions and could be incorporated in the benchmark workflow for segmentation and TMTV calculation in DLBCL (8). The main limitation of LIONZ SUV4 is the segmentation of healthy high uptake organs or areas when located nearby tumor regions. This is a consequence of using SUV4.0 in the pipeline. However, this approach eliminates LIONZ over-segmentation issue which occurred in almost every patient as we see in Fig. 4 A and in the Bland-Altman plot (Supplemental Fig. 1A). In largely diffused tumors, LIONZ failed to identify the tumor borders and leads to a large underestimation of the tumor region. This is also fixed with LIONZ SUV4 (Fig. 2 ). Even though LIONZ SUV4 pipeline requires reader interaction, these are minimized compared to SUV4.0 segmentation pipeline and still achieved reliable segmentations with equivalent performance when used in a prognostic model. Moreover, the coefficients of the prediction model were not recalibrated for LIONZ and LIONZ SUV4 based PET parameters. The coefficients from the model derived using SUV4.0 PET parameters, as originally described in (5), have been used in this study. The PET based parameter values from the new segmentations were simply replaced by those from SUV4.0 segmentation to obtain the new TTP probabilities. The LIONZ tool was developed by the ENHANCE PET group at the Medical University of Vienna and was trained on the open-source AutoPET Challenge dataset, which includes over 1000 18 F-FDG PET studies. These studies cover patients with malignant melanoma, lymphoma, or lung cancer, as well as control subjects without cancer. Therefore, LIONZ was designed with the idea of being used not only for lymphoma but for a variety of different tumors. LIONZ and its combination with SUV4.0 threshold has also been used for the delineation of breast cancer lesions (14). Although its segmentation performance was not directly assessed, the derived segmentations were used for radiomics analyses with the aim of improving patient stratification. Droguet et al . (15) investigated the potential of LIONZ for lesion segmentation in patients with non-small cell lung cancer and also used SUV4.0 threshold to refine the delineation of the lesions. Their findings align with those of our study showing that the combination of LIONZ and SUV4.0 is a powerful tool for lesion identification and delineation. Moreover, they found that the use of LIONZ improved inter and intra-reader reproducibility. Our study suggests that this also applies for segmentation of lesions in FDG PET images of DLBCL patients when using this approach. In this paper, we selected SUV4.0 as our reference method since Barrington et al . (6) found it to be the most robust method for segmentation in DLBCL. In addition, this method was recently proposed as an international lymphoma MTV benchmark segmentation approach (8). Moreover, the Clinical PET model was initially developed using SUV4.0 segmentations. When replacing the parameter values by those derived from LIONZ SUV4 segmentations, model performance remained statistically equivalent to the original model performance. CONCLUSION LIONZ SUV4 segmentations showed a high spatial overlap with SUV4.0 segmentations and the PET parameters extracted using segmentations from these two methods were strongly correlated. The prediction model for prediction of 2y TTP using PET and clinical parameters (MTV, SUVpeak and Dmaxbulk, age and WHO status) derived from LIONZ SUV4 segmentations performed equivalent to the model which used PET parameters from SUV4.0 segmentations. Furthermore, LIONZ SUV4 required fewer reader interactions compared to the benchmark SUV4.0 segmentation workflow. These findings suggest that LIONZ SUV4 is a suitable method for tumor segmentation in DLBCL PET images and is an alternative to threshold only based segmentation methods and use of LIONZ SUV4 potentially leads to a decreased reader-variability. Abbreviations Diffuse Large B-Cell Lymphoma (DLBCL) 18 F-Fluorodeoxyglucose ( 18 F-FDG) Positron Emission Tomography (PET) Computed Tomography (CT) Metabolic Tumor Volume (MTV) Maximum distance between the largest lesion and any other lesion (Dmaxbulk) Standardized Uptake Value (SUV) Artificial Intelligence (AI) World Health Organization (WHO) Time to Progression (TTP) Quality Control (QC) Dice Similarity Coefficient (DSC) Area Under the Curve (AUC) Interquartile Range (IQR) Declarations 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. 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, S.C.A.V., J.J.E., S.E.W., G.J.C.Z., M.W.H. and R.B. declare no competing financial interests. P.J.L. received research funding from Takeda, Servier and Roche and received honoraria for advisory boards from Takeda, Servier, Genentech, Genmab, Celgene, Incyte and AbbVie. 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. Funding . This work was financially supported by the Hanarth Fonds Fund and the Dutch Cancer Society (#VU-2018-11648). The sponsor had no role in gathering, analyzing or interpreting the data. Authors contribution. This is a multidisciplinary study where different datasets were used. This implied the collaboration of multiple authors gathering a technical, clinical and international team. S.S.V.G., R.B. and M.C.F. contributed to the concept and design of the study. P.J.L. and G.J.C.Z. were responsible for acquiring and collecting the data. J.J.E., S.E.W., S.C.A.V. and M.C.F. performed the data analysis. M.C.F. completed the first draft of the manuscript. All authors reviewed and approved the manuscript. Acknowledgements . This work was financially supported by the Hanarth Fonds Fund and the Dutch Cancer Society (#VU-2018-11648). The sponsor had no role in gathering, analyzing or interpreting the data. The authors thank all the patients who participated in the trials. Ethics Declaration . All individual participants included in the study gave written informed consent to participate in the study. The HOVON-84 study was approved by the institutional review board of the Erasmus MC (2007–055) and was performed in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Consent to participate . Not applicable. References Sehn LH, Donaldson J, Chhanabhai M, Fitzgerald C, Gill K, Klasa R, et al. Introduction of combined CHOP plus rituximab therapy dramatically improved outcome of diffuse large B-cell lymphoma in British Columbia. J Clin Oncol. 2005;23(22):5027-33. Shipp M. A predictive model for aggressive non-Hodgkin's lymphoma. N Engl J Med. 1993;329(14):987-94. Zhou Z, Sehn LH, Rademaker AW, Gordon LI, Lacasce AS, Crosby-Thompson A, et al. An enhanced International Prognostic Index (NCCN-IPI) for patients with diffuse large B-cell lymphoma treated in the rituximab era. Blood. 2014;123(6):837-42. Mikhaeel NG, Heymans MW, Eertink JJ, de Vet HCW, Boellaard R, Duhrsen U, et al. Proposed New Dynamic Prognostic Index for Diffuse Large B-Cell Lymphoma: International Metabolic Prognostic Index. J Clin Oncol. 2022;40(21):2352-60. 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. Barrington SF, Zwezerijnen B, de Vet HCW, Heymans MW, Mikhaeel NG, Burggraaff CN, et al. Automated Segmentation of Baseline Metabolic Total Tumor Burden in Diffuse Large B-Cell Lymphoma: Which Method Is Most Successful? A Study on Behalf of the PETRA Consortium. J Nucl Med. 2021;62(3):332-7. Lalith Shiyam MP, Sebastian Gutschmayer. LalithShiyam/LION: lionz-v.0.9.1 (lionz-v.0.9.1) Zenodo2024 [ Boellaard R, Buvat I, Nioche C, Ceriani L, Cottereau AS, Guerra L, et al. International Benchmark for Total Metabolic Tumor Volume Measurement in Baseline (18)F-FDG PET/CT of Lymphoma Patients: A Milestone Toward Clinical Implementation. J Nucl Med. 2024;64(1):83. 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. 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. Boellaard R. Quantitative oncology molecular analysis suite: ACCURATE. Journal of Nuclear Medicine. 2018;59(1):1753. Delong ER, Delong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics. 1988;33(837-845). 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. Julie A, Droguet M, Kumar Shiyam Sundar L, Seban RD, Luporsi M, Pires M, et al. Prognostic stratification of metastatic triple-negative breast cancer patients using PET-radiomic features from malignant and tumor-free regions. Journal of Nuclear Medicine. 2024; 65 241952. Droguet M, Kumar Shiyam Sundar L, Manuel Pires, Hovhannisyan Baghdasarian N, Captier N, Luporsi M, et al. Automated segmentation of lesions in [18F]FDG PET/CT images of lung cancer patients: external evaluation of an AI-driven lesion segmentation tool (LION). Journal of Nuclear Medicine. 2024;65:241927. Tables Table 1 . Summary of statistics for the PET parameters for all three methods. SUV4.0 Median (IQR) LIONZ Median (IQR) LIONZ SUV4 Median (IQR) MTV (ml) 385.9 (133.9, 923.2) 459.1 (177.5, 1151.2) 378 (124.3, 907.8) SUVpeak 17.6 (13.4, 23.5) 18.3 (13.8, 24.2) 18.3 (13.8, 24.2) Dmaxbulk (mm) 318.8 (142.8, 468.2) 347.6 (198.1, 485.8) 330.9 (158.1, 459.1) Abbreviations. MTV: metabolic tumor volume, IQR: interquartile range Table 2. Risk classification of patients based on the Clinical PET model for LIONZ prediction model. High risk Intermediate risk Low risk Total # of patients per group for SUV4.0 High risk 48 11 0 59 Intermediate risk 11 72 7 90 Low risk 2 13 132 147 Total # of patients per group for LIONZ 61 96 139 TOTAL # of patients = 296 Table 3. Risk classification of patients based on the Clinical PET model for LIONZ SUV4 prediction model. High risk Intermediate risk Low risk Total # of patients per group for SUV4.0 High risk 44 15 0 59 Intermediate risk 1 74 15 90 Low risk 0 2 145 147 Total # of patients per group for LIONZ SUV4 45 91 160 TOTAL # of patients = 296 Additional Declarations The authors declare potential competing interests as follows: 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, S.C.A.V., J.J.E., S.E.W., G.J.C.Z., M.W.H. and R.B. declare no competing financial interests. P.J.L. received research funding from Takeda, Servier and Roche and received honoraria for advisory boards from Takeda, Servier, Genentech, Genmab, Celgene, Incyte and AbbVie. 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. Supplementary Files Supplementalfigure1.docx Supplementalfigure2.docx Supplementalfigure3.docx Supplementalfigure4.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6294601","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":433120260,"identity":"91e4ae54-50d4-49a1-aaf4-d581794f4797","order_by":0,"name":"Maria C. Ferrández","email":"data:image/png;base64,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","orcid":"https://orcid.org/0009-0007-6084-5266","institution":"Cancer Center Amsterdam","correspondingAuthor":true,"prefix":"","firstName":"Maria","middleName":"C.","lastName":"Ferrández","suffix":""},{"id":433177738,"identity":"21d2cf4f-365d-4bcf-b74a-ee87d8d79b65","order_by":1,"name":"Sandeep S. V. Golla","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Sandeep","middleName":"S. V.","lastName":"Golla","suffix":""},{"id":433177739,"identity":"81b33dd8-790f-4292-8803-416cf27d0f09","order_by":2,"name":"Sara C. A. De Visser","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Sara","middleName":"C. A.","lastName":"De Visser","suffix":""},{"id":433177740,"identity":"9aaa0eff-e77a-4409-9c3c-1f32aa2b656c","order_by":3,"name":"Jakoba J. Eertink","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Jakoba","middleName":"J.","lastName":"Eertink","suffix":""},{"id":433177741,"identity":"24d95b7f-91d2-454a-8f8e-a1f84032ddf9","order_by":4,"name":"Pieternella J. Lugtenburg","email":"","orcid":"","institution":"Erasmus MC Cancer Institute","correspondingAuthor":false,"prefix":"","firstName":"Pieternella","middleName":"J.","lastName":"Lugtenburg","suffix":""},{"id":433177742,"identity":"5d97ca55-7f6a-49de-88db-53d6d3816481","order_by":5,"name":"Sanne E. Wiegers","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Sanne","middleName":"E.","lastName":"Wiegers","suffix":""},{"id":433177743,"identity":"2b5d8740-f400-477e-a3fd-e912fb16d432","order_by":6,"name":"Gerben J. C. Zwezerijnen","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Gerben","middleName":"J. C.","lastName":"Zwezerijnen","suffix":""},{"id":433177744,"identity":"0a5c12d9-c81f-478a-aaf2-5416873916e2","order_by":7,"name":"Martijn W. Heymans","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Martijn","middleName":"W.","lastName":"Heymans","suffix":""},{"id":433177745,"identity":"5d6abe8d-f7d3-4d43-b60b-b15864896cf1","order_by":8,"name":"Josée M. Zijlstra","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Josée","middleName":"M.","lastName":"Zijlstra","suffix":""},{"id":433177746,"identity":"707cbb9c-2806-4dbc-b342-1a60790676db","order_by":9,"name":"Ronald Boellaard","email":"","orcid":"","institution":"Cancer Center Amsterdam","correspondingAuthor":false,"prefix":"","firstName":"Ronald","middleName":"","lastName":"Boellaard","suffix":""}],"badges":[],"createdAt":"2025-03-24 11:07:20","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":true,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":true,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6294601/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6294601/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":79265458,"identity":"eb2c9c91-24e7-40e9-9292-af5f4a0ade5c","added_by":"auto","created_at":"2025-03-26 10:01:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":352202,"visible":true,"origin":"","legend":"\u003cp\u003eOversegmentation of lesions. (A) Example of patient where LIONZ oversegments lesions and the corresponding corrected segmentation performed by LIONZ\u003csup\u003eSUV4\u003c/sup\u003e with the MTV values (mL). (B) Axial, coronal and sagittal views of lesion from example patient with the LIONZ segmentation contour in red and the LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentation contour in green.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/d2f99a7eae26a33ffe54191f.png"},{"id":79263771,"identity":"1cdfe090-cb32-4ad0-bc9b-f2cd7f53db1f","added_by":"auto","created_at":"2025-03-26 09:53:09","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":400879,"visible":true,"origin":"","legend":"\u003cp\u003eThree examples of undersegmentation of lesions by LIONZ and the corresponding MTV values (mL). Undersegmentation was generally corrected after using the combined approach LIONZ\u003csup\u003eSUV4\u003c/sup\u003e.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/212dd3090d9b3aa79c09ba6b.png"},{"id":79263762,"identity":"6ee2809d-8b4d-47ab-a4b6-c23ed3788ab2","added_by":"auto","created_at":"2025-03-26 09:53:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":368481,"visible":true,"origin":"","legend":"\u003cp\u003eExamples of manual corrections for LIONZ\u003csup\u003eSUV4\u003c/sup\u003e. From left to right, an injection spot, a kidney and an object outside the body.\u0026nbsp;\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/61d0393c09bfd6a76dbb3c29.png"},{"id":79263748,"identity":"16d8fdfc-94e3-4d41-a0d5-ba297f577328","added_by":"auto","created_at":"2025-03-26 09:53:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":78246,"visible":true,"origin":"","legend":"\u003cp\u003eAnalysis of MTV values (A) Pearson correlation of MTV derived from LIONZ segmentation and reference segmentation method (SUV4.0) (B) Pearson correlation of MTV derived from LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentation and reference segmentation method (SUV4.0). LOI: line of identity.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/8a1a3f5f3202a4635a98257c.png"},{"id":79265460,"identity":"6f27fa87-86dc-4961-b203-cca81b60fc81","added_by":"auto","created_at":"2025-03-26 10:01:08","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":66006,"visible":true,"origin":"","legend":"\u003cp\u003eAnalysis of SUVpeak values (A) Pearson correlation of SUVpeak derived from LIONZ segmentation and reference segmentation method (SUV4.0) (B) Pearson correlation of SUVpeak derived from LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentation and reference segmentation method (SUV4.0). LOI: line of identity.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/9643dcfd7eac743520c9eaba.png"},{"id":79263766,"identity":"89780d1f-931b-4c66-97e2-f515afcf7c53","added_by":"auto","created_at":"2025-03-26 09:53:09","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":83324,"visible":true,"origin":"","legend":"\u003cp\u003eAnalysis of Dmaxbulk values (A) Pearson correlation of Dmaxbulk derived from LIONZ segmentation and reference segmentation method (SUV4.0) (B) Pearson correlation of Dmaxbulk derived from LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentation and reference segmentation method (SUV4.0). LOI: line of identity.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/f637b921fc0412b1d118cfc0.png"},{"id":79265467,"identity":"f3b73e62-00f5-444d-985b-6c0f77cb1c7c","added_by":"auto","created_at":"2025-03-26 10:01:10","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":104994,"visible":true,"origin":"","legend":"\u003cp\u003eAnalysis of TTP probabilities (A) Pearson correlation of probabilities derived from LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentation and reference segmentation method (SUV4.0) (B) Area under the curve for prediction models using LIONZ\u003csup\u003eSUV4\u003c/sup\u003e and SUV4.0 segmentation method. LOI: line of identity.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/d77493366167d56f782b6bef.png"},{"id":79265478,"identity":"214ec905-133b-4b5e-ad56-f579dfa43c36","added_by":"auto","created_at":"2025-03-26 10:01:15","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2182296,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/e287dc72-248f-499c-8399-d9c256d32aaf.pdf"},{"id":79265459,"identity":"c1b7d2e7-0a55-438c-aa42-93274b87338a","added_by":"auto","created_at":"2025-03-26 10:01:08","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":166320,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalfigure1.docx","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/0dfa76f12b5a3ecc74455436.docx"},{"id":79263747,"identity":"fc7193f4-b1e9-4ea4-9155-4c0cde161dbb","added_by":"auto","created_at":"2025-03-26 09:53:08","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":148942,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalfigure2.docx","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/9fada3afd7b4edce1e66f7d8.docx"},{"id":79263760,"identity":"16b8b787-9483-482c-bb07-f2a6a202c5da","added_by":"auto","created_at":"2025-03-26 09:53:09","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":165001,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalfigure3.docx","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/240441d4c9a13951f8c079c9.docx"},{"id":79263778,"identity":"27c72dde-8e5e-4451-bdb6-85499a25ca5f","added_by":"auto","created_at":"2025-03-26 09:53:09","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":169266,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalfigure4.docx","url":"https://assets-eu.researchsquare.com/files/rs-6294601/v1/1c24366ac8e5a0272da241cb.docx"}],"financialInterests":"The authors declare potential competing interests as follows: 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, S.C.A.V., J.J.E., S.E.W., G.J.C.Z., M.W.H. and R.B. declare no competing financial interests. P.J.L. received research funding from Takeda, Servier and Roche and received honoraria for advisory boards from Takeda, Servier, Genentech, Genmab, Celgene, Incyte and AbbVie. 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.","formattedTitle":"\u003cp\u003eEvaluation of an artificial intelligence method for lesion segmentation of baseline FDG PET studies of DLBCL patients\u003c/p\u003e","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eDiffuse large B cell lymphoma (DLBCL) is the most common type of non-Hodgkin lymphoma, accounting for 30% of all non-Hodgkin lymphoma diagnoses (1). It is characterized by diffused patterns of growth and for being an extremely heterogeneous disease, which hinders treatment planning. The International Prognostic Index (IPI), and variations of these parameters, have served clinicians in risk stratification and guided treatment strategies for DLBCL patients (2\u0026ndash;4). However, advancements in diagnostics and therapeutic treatments have significantly improved DLBCL outcomes particularly for high-risk patients (1). As a result, IPI\u0026rsquo;s effectiveness in reliably predicting treatment failure has declined (4).\u003c/p\u003e \u003cp\u003e \u003csup\u003e18\u003c/sup\u003eF-Fluorodeoxyglucose (\u003csup\u003e18\u003c/sup\u003eF-FDG) positron emission tomography (PET) in combination with computed tomography (CT) is a standard tool for the prognostic evaluation of DLBCL. There is evidence that parameters quantified through \u003csup\u003e18\u003c/sup\u003eF-FDG PET/CT imaging have predictive potential in lymphoma, such as the metabolic tumor volume (MTV) and the maximum distance between the bulkiest lesion and any other lesion (Dmaxbulk) (5). Segmentation of tumors is required to quantify such parameters from the PET images. Standard uptake value (SUV) thresholding methods are most commonly used for tumor segmentation. In the case of DLBCL, SUV4.0 is the segmentation method which has shown to be most robust (6). However, these semi-automatic methods require manual input and/or lesion selection thereby becoming time-consuming and sensitive to reader variability.\u003c/p\u003e \u003cp\u003eIn this study we used LIONZ (7), an externally developed artificial intelligence (AI) tool, to automatically detect DLBCL lesions. We further optimized LIONZ by combining it with SUV4.0 thresholding (LIONZ\u003csup\u003eSUV4\u003c/sup\u003e). The segmentations derived from LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e were compared to those of SUV4.0. SUV4.0 has been referred to as the preferred segmentation threshold in DLBCL, as shown by Barrington \u003cem\u003eet al\u003c/em\u003e. (6) and by the international lymphoma MTV benchmark (8). The benchmark established that small lesions with a volume lower than 3 mL needs to be manually included in the SUV4.0 segmentation. Additionally, MTV, SUVpeak and Dmaxbulk values were extracted and analysed for all the three segmentation methods. Furthermore, we assessed the predictive power of PET parameters derived from the LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations using the prediction model developed by Eertink \u003cem\u003eet al.\u003c/em\u003e (5). The model is a logistic regression outcome prediction model which utilizes MTV, Dmaxbulk and SUVpeak parameters extracted from SUV4.0 segmentations together with two clinical parameters, age and World Health Organization (WHO) performance status. The model outcome is a dichotomous value equivalent to the probability of time to progression (TTP) within or longer than 2 years in DLBCL patients.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData\u003c/h2\u003e \u003cp\u003eA total of 317 baseline DLBCL \u003csup\u003e18\u003c/sup\u003eF-FDG PET/CT scans from the HOVON-84 trial were available for this study (EudraCT, 2006-005, 174\u0026thinsp;\u0026minus;\u0026thinsp;42) (9). This study was approved by institutional review boards and all included patients provided informed consent. The use of all data within the HOVON-84 database has been approved by the institutional review board of the VU University Medical Center (JR /20140414). From the 317 patients, 7 were lost to follow-up within 2 years and 14 other patients died within 2 years of unrelated reasons. This led to a total of 296 DLBCL patients included in this study. Quality control of scans was conducted following the criteria set by the European Association of Nuclear Medicine (10) and is fully described elsewhere (5). The scans included in this study are compliant with EARL1 standards.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSUV4.0 segmentation\u003c/h3\u003e\n\u003cp\u003eLesions were delineated through an automated preselection defined by a SUV\u0026thinsp;\u0026ge;\u0026thinsp;4.0 and a volume threshold of \u0026ge;\u0026thinsp;3 mL. Non-tumor regions were removed, and lesions smaller than 3 mL were interactively added with one-click per lesion. If tumor regions were adjacent to high intensity non-tumor regions (e.g., kidney, bladder), these were manually corrected. Further details on the delineation methods and workflow are provided in the international lymphoma MTV benchmark for segmentation (8). All scans were reviewed by a nuclear medicine physician (GJCZ), and delineations were performed under his supervision.\u003c/p\u003e\n\u003ch3\u003eAI lesion detection and segmentation\u003c/h3\u003e\n\u003cp\u003eLIONZ is a deep learning tool developed by the ENHANCE PET (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://enhance.pet/\u003c/span\u003e\u003cspan address=\"https://enhance.pet/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) group at Medical University of Vienna, Austria. LIONZ was trained on over 1000 \u003csup\u003e18\u003c/sup\u003eF-FDG PET/CT scans of patients with multiple cancer types (lymphoma, melanoma and lung cancer) and healthy subjects. LIONZ is open-source and is available for download at Zenodo (7).\u003c/p\u003e \u003cp\u003eBesides analyzing the segmentation performance of LIONZ, we developed LIONZ\u003csup\u003eSUV4\u003c/sup\u003e which combines LIONZ and SUV4.0 into the segmentation workflow for the same purpose. This new method uses LIONZ as a lesion detection and selection step and is followed by SUV4.0 threshold region growing. In a second step, one-click corrections are used to manually remove injection spots, spots outside of the body and any segmented high uptake organ adjacent to tumor regions (brain, bladder and/or kidney). These corrections are limited to removal of regions and no missed lesions were manually added unlike for the SUV4.0 method that require reader selection of lesions smaller than 3 mL (8).\u003c/p\u003e \u003cp\u003eSUV4.0, LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e methods were implemented or embedded using the same software tool ACCURATE (11). ACCURATE enables the quantitative analysis of the PET scans and computation of PET parameters. We used ACCURATE to calculate MTV, SUVpeak and Dmaxbulk for each PET scan.\u003c/p\u003e\n\u003ch3\u003ePrediction model\u003c/h3\u003e\n\u003cp\u003eThe prediction model was first developed using the HOVON-84 trial by \u003cem\u003eEertink et al.\u003c/em\u003e (5). The following features are included in this model: MTV, SUVpeak, Dmaxbulk, age and WHO performance status. These features were used as predictors in a logistic regression model to predict the probability of 2-year TTP for each patient. For the definition of TTP, patients who died within 2 years starting at the time of baseline scan without signs of progression were excluded from the analysis.\u003c/p\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eThe dice similarity coefficient (DSC) was used to compare the performance of the three different segmentations. Pearson correlation and Bland-Altman plots were used to assess the relation between segmentation methods for all three PET parameters: MTV, SUVpeak and Dmaxbulk values. In order to assess the predictive potential the PET parameters derived from the LIONZ and LIONZ \u003csup\u003eSUV4\u003c/sup\u003e segmentations, we replaced the features\u0026rsquo; values in the logistic regression of the Clinical PET model by those of the LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e based MTV, SUVpeak and Dmaxbulk values. We used the area under the curve (AUC) to assess the model\u0026rsquo;s performance. The Delong test was used to compare AUC values between methods (12). In addition, a risk classification analysis of the predicted TTP values yielded by the prediction model using the three different segmentations was also performed. Using the predicted TTP values derived from the model using SUV4.0 segmentations, 20% of patients with the highest predicted values were classified as high risk, 50% of the patients with the lowest predicted values in the low risk group and the rest as the intermediate risk group. The risk classifications from the prediction model using SUV4.0 segmentations was used as reference and, the effect on the risk classifications when using the predicted TTP values from other segmentations (LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e) was evaluated.\u003c/p\u003e \u003c/div\u003e"},{"header":"RESULTS","content":"\u003cp\u003eEarly in the analysis, we found that LIONZ consistently identified areas beyond the actual lesion borders and labeling surrounding healthy tissue areas as part of the lesion. This led to an overestimation of the tumor size and volume (i.e. over-segmentation). Moreover, in cases with largely disseminated tumors and smooth tumor borders, LIONZ tended to under-segment lesions leaving out regions of interest that should be included in the segmentation (i.e. under-segmentation). Therefore, combining LIONZ with a SUV4.0 threshold (LIONZ\u003csup\u003eSUV4\u003c/sup\u003e) was evaluated as an option to correct the under/over-segmentation by LIONZ. In the LIONZ\u003csup\u003eSUV4\u003c/sup\u003e method, LIONZ is used as a lesion detection step, followed by SUV4.0 threshold based region shrinking/growing, which compensates for any over/under-segmentation from LIONZ. Some examples that illustrate how LIONZ\u003csup\u003eSUV4\u003c/sup\u003e led to a more accurate segmentation compared to LIONZ are given in Figures 1 and 2. Occasionally selected high uptake organs such as brain, bladder and/or kidneys adjacent to tumor regions were also manually removed for both LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e. The most complicated cases involved removing the kidneys slice by slice from the PET images but these accounted for only 7 cases in total (2%). Examples of these corrections are given in Figure 4. Unlike the corrections for the SUV4.0 segmentation, no lesions were manually added.\u003c/p\u003e\n\u003cp\u003eThe DSC was calculated for all segmentations from LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e with SUV4.0 as the reference segmentation. The median DSC and interquartile range (IQR) resulted in 0.77 (0.64 - 0.84) for LIONZ and 0.87 (0.80 - 0.93) for LIONZ\u003csup\u003eSUV4\u003c/sup\u003e. There were 6 segmentations for which the DSC was equal to 0. These 6 cases corresponded to very small lesions (\u0026lt;3ml) in the SUV4.0 segmentations. For 4 of these cases, LIONZ failed in detecting any lesions. Unlike for LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e, when using SUV4.0 segmentation, small lesions (\u0026lt;3ml) are manually added as indicated in (8).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMTV values were highly correlated for LIONZ and SUV4.0 (0.80, p\u0026lt;0.0001). Generally, LIONZ overestimates tumor volume compared to SUV4.0, as we see in Figure 4 and Supplemental Figure 1A. The cases were LIONZ underestimates MTV correspond to those patients with largely diffused and non-defined tumor regions. When using LIONZ\u003csup\u003eSUV4\u003c/sup\u003e we found that MTV values were more aligned with those of SUV4.0, with a higher correlation (0.98, p\u0026lt;0.0001). Moreover, the Bland-Altman plot in Supplemental Figure 1B shows an improved agreement of the MTV values when using LIONZ\u003csup\u003eSUV4\u003c/sup\u003e compared to LIONZ.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRegarding SUVpeak, the correlation was similar to that of MTV (0.98, p-value \u0026lt;0.0001). This was the case for both LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e. LIONZ and LIONZ\u003csup\u003eSUV4\u0026nbsp;\u003c/sup\u003eSUVpeak values which led to 0 correspond to patients in which tumors were not detected at all (Figure 5). A slight overestimation of SUVpeak values is illustrated in the Bland-Altman plots for both LIONZ and \u0026nbsp;LIONZ\u003csup\u003eSUV4\u003c/sup\u003e (Supplemental Figure 2).\u003c/p\u003e\n\u003cp\u003ePearson correlation for Dmaxbulk was of 0.83 (p-value \u0026lt; 0.0001) for LIONZ and SUV4.0 values. For multiple patients, LIONZ identified small lesions that do not appear in the SUV4.0 segmentations which leads to large differences in Dmaxbulk (Figure 6). This seems to be improved when using LIONZ\u003csup\u003eSUV4\u003c/sup\u003e (Supplemental Figure 3). The correlation was slightly higher when comparing LIONZ\u003csup\u003eSUV4\u0026nbsp;\u003c/sup\u003eand SUV4.0 (0.89, p\u0026lt;0.0001). Cases with the largest differences were usually patients with multiple small lesions largely spread. A summary of statistics for these parameters is given in Table 1.\u003c/p\u003e\n\u003cp\u003eTTP probabilities were calculated with the prediction model using both the LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations and compared to those of the original model (i.e. using SUV4.0 segmentations). The Pearson correlation was of 0.90 and 0.96 respectively (p-value \u0026lt; 0.0001, Figure 7A). The agreement between probabilities increased when using LIONZ\u003csup\u003eSUV4\u003c/sup\u003e compared to LIONZ segmentations as shown in the Bland-Altman plot in Supplemental Figure 4. The AUC for the prediction model derived from the SUV4.0 segmentations is 0.79. Replacing the SUV4.0 parameters with those derived from the LIONZ segmentations yielded an AUC of 0.74 and for LIONZ\u003csup\u003eSUV4\u003c/sup\u003e, the model achieved an AUC of 0.78 (Figure 7B). The Delong test found the AUC values for the model derived from LIONZ segmentations and the model derived from SUV4.0 segmentations to be statistically different (P\u0026lt;0.0001). No statistically significant difference was found between the AUC curves of the model derived from LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations and the model derived from the SUV4.0 segmentation (P \u0026gt; 0.05). In Tables 2 and 3 we illustrate the changes in the risk classification from the model using SUV4.0 segmentations to the models using LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations. Regarding the model that used LIONZ segmentations, the number of patients that remained in the same risk groups as the model derived from the SUV4.0 segmentation was 48 out of 59, 72 out of 90 and 132 out of 147 for the high, intermediate, and low risk group respectively. In the case of LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentation, 44 out of 59, 74 out of 90, and 145 out of 147 remained in the high, intermediate and low risk groups respectively.\u0026nbsp;\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eAI based models are quickly advancing in the PET imaging field with multiple applications. Automatic segmentation is one of the most promising tasks (13). However, segmentation of lymphoma can be challenging due to the heterogeneity of the disease. In this study we investigated the potential of LIONZ, an AI method for the detection of DLBCL lesions. Moreover, we developed LIONZ\u003csup\u003eSUV4\u003c/sup\u003e, which combined LIONZ and SUV4.0 in an effort to overcome LIONZ tendency to over/under-segment DLBCL lesions. We found that the segmentations yielded by LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e exhibited a high level of similarity with those of SUV4.0 with a median DSC of 0.77 and 0.87, respectively. Moreover, the PET extracted parameters derived from LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations are highly correlated with those of SUV4.0. These findings suggest that LIONZ itself is capable of identifying suspicious uptake sites and accurately discarding physiological regions with high uptake for nearly all patients. The modified version including SUV4.0 (LIONZ\u003csup\u003eSUV4\u003c/sup\u003e) is meant to compensate for the over-/under segmentation tendency of LIONZ, as shown in Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Overall, LIONZ\u003csup\u003eSUV4\u003c/sup\u003e shows a stronger agreement of PET parameters with SUV4.0. Manual corrections were added to the pipeline to remove non-tumor regions. These corrections are mostly one-click interactions. In general, LIONZ accurately identified smaller lesions (\u0026lt;\u0026thinsp;3 ml) in the PET images, therefore we did not include any corrections involving manually adding suspected missed lesions, unlike for SUV4.0 corrections where many lesions would have been missed and therefore readers are required to manually add missed lesions as indicated in (8). The manual addition of missed lesions inevitably increases the number of reader interactions which is one of the reasons why SUV4.0 segmentations suffer from inter-reader variability. The new approach proposed by LIONZ\u003csup\u003eSUV4\u003c/sup\u003e, requires fewer interactions by decreasing the discrepancies regarding non-detected smaller lesions (\u0026lt;\u0026thinsp;3ml), eventually decreasing inter-reader variability. These findings suggest that LIONZ is an efficient tool for the identification of lesions and could be incorporated in the benchmark workflow for segmentation and TMTV calculation in DLBCL (8).\u003c/p\u003e \u003cp\u003eThe main limitation of LIONZ\u003csup\u003eSUV4\u003c/sup\u003e is the segmentation of healthy high uptake organs or areas when located nearby tumor regions. This is a consequence of using SUV4.0 in the pipeline. However, this approach eliminates LIONZ over-segmentation issue which occurred in almost every patient as we see in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA and in the Bland-Altman plot (Supplemental Fig.\u0026nbsp;1A). In largely diffused tumors, LIONZ failed to identify the tumor borders and leads to a large underestimation of the tumor region. This is also fixed with LIONZ\u003csup\u003eSUV4\u003c/sup\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Even though LIONZ\u003csup\u003eSUV4\u003c/sup\u003e pipeline requires reader interaction, these are minimized compared to SUV4.0 segmentation pipeline and still achieved reliable segmentations with equivalent performance when used in a prognostic model. Moreover, the coefficients of the prediction model were not recalibrated for LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e based PET parameters. The coefficients from the model derived using SUV4.0 PET parameters, as originally described in (5), have been used in this study. The PET based parameter values from the new segmentations were simply replaced by those from SUV4.0 segmentation to obtain the new TTP probabilities.\u003c/p\u003e \u003cp\u003eThe LIONZ tool was developed by the ENHANCE PET group at the Medical University of Vienna and was trained on the open-source \u003cem\u003eAutoPET Challenge\u003c/em\u003e dataset, which includes over 1000 \u003csup\u003e18\u003c/sup\u003eF-FDG PET studies. These studies cover patients with malignant melanoma, lymphoma, or lung cancer, as well as control subjects without cancer. Therefore, LIONZ was designed with the idea of being used not only for lymphoma but for a variety of different tumors. LIONZ and its combination with SUV4.0 threshold has also been used for the delineation of breast cancer lesions (14). Although its segmentation performance was not directly assessed, the derived segmentations were used for radiomics analyses with the aim of improving patient stratification. \u003cem\u003eDroguet et al\u003c/em\u003e. (15) investigated the potential of LIONZ for lesion segmentation in patients with non-small cell lung cancer and also used SUV4.0 threshold to refine the delineation of the lesions. Their findings align with those of our study showing that the combination of LIONZ and SUV4.0 is a powerful tool for lesion identification and delineation. Moreover, they found that the use of LIONZ improved inter and intra-reader reproducibility. Our study suggests that this also applies for segmentation of lesions in FDG PET images of DLBCL patients when using this approach.\u003c/p\u003e \u003cp\u003eIn this paper, we selected SUV4.0 as our reference method since Barrington \u003cem\u003eet al\u003c/em\u003e. (6) found it to be the most robust method for segmentation in DLBCL. In addition, this method was recently proposed as an international lymphoma MTV benchmark segmentation approach (8). Moreover, the Clinical PET model was initially developed using SUV4.0 segmentations. When replacing the parameter values by those derived from LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations, model performance remained statistically equivalent to the original model performance.\u003c/p\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eLIONZ\u003csup\u003eSUV4\u0026nbsp;\u003c/sup\u003esegmentations showed a high spatial overlap with SUV4.0 segmentations and the PET parameters extracted using segmentations from these two methods were strongly correlated. The prediction model for prediction of 2y TTP using PET and clinical parameters (MTV, SUVpeak and Dmaxbulk, age and WHO status) derived from LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations performed equivalent to the model which used PET parameters from SUV4.0 segmentations. Furthermore, LIONZ\u003csup\u003eSUV4\u003c/sup\u003e required fewer reader interactions compared to the benchmark SUV4.0 segmentation workflow. These findings suggest that LIONZ\u003csup\u003eSUV4\u0026nbsp;\u003c/sup\u003eis a suitable method for tumor segmentation in DLBCL PET images and is an alternative to threshold only based segmentation methods and use of LIONZ\u003csup\u003eSUV4\u003c/sup\u003e potentially leads to a decreased reader-variability.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eDiffuse Large B-Cell Lymphoma (DLBCL)\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e18\u003c/sup\u003eF-Fluorodeoxyglucose (\u003csup\u003e18\u003c/sup\u003eF-FDG)\u003c/p\u003e\n\u003cp\u003ePositron Emission Tomography (PET)\u003c/p\u003e\n\u003cp\u003eComputed Tomography (CT)\u003c/p\u003e\n\u003cp\u003eMetabolic Tumor Volume (MTV)\u003c/p\u003e\n\u003cp\u003eMaximum distance between the largest lesion and any other lesion (Dmaxbulk)\u003c/p\u003e\n\u003cp\u003eStandardized Uptake Value (SUV)\u003c/p\u003e\n\u003cp\u003eArtificial Intelligence (AI)\u003c/p\u003e\n\u003cp\u003eWorld Health Organization (WHO)\u003c/p\u003e\n\u003cp\u003eTime to Progression (TTP)\u003c/p\u003e\n\u003cp\u003eQuality Control (QC)\u003c/p\u003e\n\u003cp\u003eDice Similarity Coefficient (DSC)\u003c/p\u003e\n\u003cp\u003eArea Under the Curve (AUC)\u003c/p\u003e\n\u003cp\u003eInterquartile Range (IQR)\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAvailability of data and materials.\u0026nbsp;\u003c/strong\u003eThe datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e. 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, S.C.A.V., J.J.E., S.E.W., G.J.C.Z., M.W.H. and R.B. declare no competing financial interests. P.J.L. received research funding from Takeda, Servier and Roche and received honoraria for advisory boards from Takeda, Servier, Genentech, Genmab, Celgene, Incyte and AbbVie. J.M.Z. received research funding from Roche and received honoraria for advisory boards from\u0026nbsp;Takeda, Gilead, BMS and Roche. No other potential conflicts of interest relevant to this article exist.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e. This work was financially supported by the Hanarth Fonds Fund and the Dutch Cancer Society (#VU-2018-11648). The sponsor had no role in gathering, analyzing or interpreting the data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors contribution.\u0026nbsp;\u003c/strong\u003eThis is a multidisciplinary study where different datasets were used. This implied the collaboration of multiple authors gathering a technical, clinical and international team. S.S.V.G., R.B. and M.C.F. contributed to the concept and design of the study. P.J.L. and G.J.C.Z. were responsible for acquiring and collecting the data. J.J.E., S.E.W., S.C.A.V. and M.C.F. performed the data analysis. M.C.F. completed the first draft of the manuscript. All authors reviewed and approved the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e. This work was financially supported by the Hanarth Fonds Fund and the Dutch Cancer Society (#VU-2018-11648). The sponsor had no role in gathering, analyzing or interpreting the data. The authors thank all the patients who participated in the trials.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Declaration\u003c/strong\u003e. All individual participants included in the study gave written informed consent to participate in the study. The HOVON-84 study was approved by the institutional review board of the Erasmus MC (2007–055) and was performed in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to participate\u003c/strong\u003e. Not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSehn LH, Donaldson J, Chhanabhai M, Fitzgerald C, Gill K, Klasa R, et al. Introduction of combined CHOP plus rituximab therapy dramatically improved outcome of diffuse large B-cell lymphoma in British Columbia. J Clin Oncol. 2005;23(22):5027-33.\u003c/li\u003e\n\u003cli\u003eShipp M. A predictive model for aggressive non-Hodgkin\u0026apos;s lymphoma. N Engl J Med. 1993;329(14):987-94.\u003c/li\u003e\n\u003cli\u003eZhou Z, Sehn LH, Rademaker AW, Gordon LI, Lacasce AS, Crosby-Thompson A, et al. An enhanced International Prognostic Index (NCCN-IPI) for patients with diffuse large B-cell lymphoma treated in the rituximab era. Blood. 2014;123(6):837-42.\u003c/li\u003e\n\u003cli\u003eMikhaeel NG, Heymans MW, Eertink JJ, de Vet HCW, Boellaard R, Duhrsen U, et al. Proposed New Dynamic Prognostic Index for Diffuse Large B-Cell Lymphoma: International Metabolic Prognostic Index. J Clin Oncol. 2022;40(21):2352-60.\u003c/li\u003e\n\u003cli\u003eEertink 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.\u003c/li\u003e\n\u003cli\u003eBarrington SF, Zwezerijnen B, de Vet HCW, Heymans MW, Mikhaeel NG, Burggraaff CN, et al. Automated Segmentation of Baseline Metabolic Total Tumor Burden in Diffuse Large B-Cell Lymphoma: Which Method Is Most Successful? A Study on Behalf of the PETRA Consortium. J Nucl Med. 2021;62(3):332-7.\u003c/li\u003e\n\u003cli\u003eLalith Shiyam MP, Sebastian Gutschmayer. LalithShiyam/LION: lionz-v.0.9.1 (lionz-v.0.9.1) Zenodo2024 [\u003c/li\u003e\n\u003cli\u003eBoellaard R, Buvat I, Nioche C, Ceriani L, Cottereau AS, Guerra L, et al. International Benchmark for Total Metabolic Tumor Volume Measurement in Baseline (18)F-FDG PET/CT of Lymphoma Patients: A Milestone Toward Clinical Implementation. J Nucl Med. 2024;64(1):83.\u003c/li\u003e\n\u003cli\u003eLugtenburg PJ, de Nully Brown P, van der Holt B, D\u0026apos;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.\u003c/li\u003e\n\u003cli\u003eBoellaard 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.\u003c/li\u003e\n\u003cli\u003eBoellaard R. Quantitative oncology molecular analysis suite: ACCURATE. Journal of Nuclear Medicine. 2018;59(1):1753.\u003c/li\u003e\n\u003cli\u003eDelong ER, Delong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics. 1988;33(837-845).\u003c/li\u003e\n\u003cli\u003eBlanc-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.\u003c/li\u003e\n\u003cli\u003eJulie A, Droguet M, Kumar Shiyam Sundar L, Seban RD, Luporsi M, Pires M, et al. Prognostic stratification of metastatic triple-negative breast cancer patients using PET-radiomic features from malignant and tumor-free regions. Journal of Nuclear Medicine. 2024; 65 241952.\u003c/li\u003e\n\u003cli\u003eDroguet M, Kumar Shiyam Sundar L, Manuel Pires, Hovhannisyan Baghdasarian N, Captier N, Luporsi M, et al. Automated segmentation of lesions in [18F]FDG PET/CT images of lung cancer patients: external evaluation of an AI-driven lesion segmentation tool (LION). Journal of Nuclear Medicine. 2024;65:241927.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e. Summary of statistics for the PET parameters for all three methods.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSUV4.0\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eMedian (IQR)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLIONZ\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eMedian (IQR)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLIONZ\u003csup\u003eSUV4\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eMedian (IQR)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMTV (ml)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 161px;\"\u003e\n \u003cp\u003e385.9 (133.9, 923.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003e459.1 (177.5, 1151.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e378 (124.3, 907.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSUVpeak\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 161px;\"\u003e\n \u003cp\u003e17.6 (13.4, 23.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003e18.3 (13.8, 24.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e18.3 (13.8, 24.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDmaxbulk (mm)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 161px;\"\u003e\n \u003cp\u003e318.8 (142.8, 468.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003e347.6 (198.1, 485.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e330.9 (158.1, 459.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations. MTV: metabolic tumor volume, IQR: interquartile range\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u003c/strong\u003e Risk classification of patients based on the Clinical PET model for LIONZ prediction model.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHigh risk\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIntermediate risk\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLow risk\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 122px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal # of patients per group for SUV4.0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHigh risk\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e48\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e59\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIntermediate risk\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e72\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLow risk\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e132\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e147\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal # of patients per group for LIONZ\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003e96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e139\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTOTAL # of patients =\u003c/strong\u003e \u003cstrong\u003e296\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u003c/strong\u003e Risk classification of patients based on the Clinical PET model for LIONZ\u003csup\u003eSUV4\u003c/sup\u003e prediction model.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 141px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHigh risk\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIntermediate risk\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLow risk\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 122px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal # of patients per group for SUV4.0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 141px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHigh risk\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e44\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e59\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 141px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIntermediate risk\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e74\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 141px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLow risk\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e145\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e147\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 141px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal # of patients per group for LIONZ\u003csup\u003eSUV4\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e160\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTOTAL # of patients =\u003c/strong\u003e \u003cstrong\u003e296\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"VU Amsterdam","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"artificial intelligence, tumor segmentation, diffuse large b cell lymphoma, PET imaging ","lastPublishedDoi":"10.21203/rs.3.rs-6294601/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6294601/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground.\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThe aim of this study is to investigate the use of an artificial intelligence (AI) method, LIONZ, in combination with an intensity-based threshold method, SUV4.0, for the automatic selection and segmentation of diffuse large B cell lymphoma (DLBCL) lymphoma lesions.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods.\u003c/b\u003e\u003c/p\u003e \u003cp\u003e296 DLBCL \u003csup\u003e18\u003c/sup\u003eF-FDG PET scans were analyzed. Metabolic tumor volume, peak standardized uptake value (SUVpeak) and, maximum distance from the bulkiest lesion to another lesion (Dmaxbulk) were extracted from the LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations and compared to those extracted from SUV4.0 segmentations using Pearson correlation (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) and Bland-Altman plots. Segmentation performance was assessed using the Dice similarity coefficient (DSC) with SUV4.0 segmentation as a reference. A prediction model which includes MTV, SUVpeak, Dmaxbulk, age and performance status was used to predict the probability of 2 year time to progression using the parameters extracted from the LIONZ, LIONZ\u003csup\u003eSUV4\u003c/sup\u003e and SUV4.0 segmentations. Association of probabilities was evaluated using Pearson correlation (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) and Bland-Altman. The area under (AUC) the curve was used to assess and compare the performance of both methods.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults.\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThe median DSC (interquartile range) for LIONZ when compared to SUV4.0 was of 0.77 (0.64\u0026ndash;0.84) and for LIONZ\u003csup\u003eSUV4\u003c/sup\u003e of 0.87 (0.80\u0026ndash;0.93). MTV, SUVpeak and Dmaxbulk from both the LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e were highly correlated to the SUV4.0 segmentations derived parameters (R\u0026thinsp;\u0026ge;\u0026thinsp;0.80, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001). LIONZ\u003csup\u003eSUV4\u003c/sup\u003e reduced overestimation of segmented areas and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e MTV showed a stronger agreement with that of SUV4.0 compared to LIONZ (0.99 and 0.80 respectively, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001). The prediction model yielded an AUC of 0.74, 0.78 and 0.79 when using segmentations from LIONZ, LIONZ\u003csup\u003eSUV4\u003c/sup\u003e and SUV4.0 respectively. The predicted probabilities yielded by the models using the LIONZ and LIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations were also highly correlated with those of SUV4.0 segmentation (0.9 and 0.96 respectively, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001).\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusion.\u003c/b\u003e\u003c/p\u003e \u003cp\u003eLIONZ\u003csup\u003eSUV4\u003c/sup\u003e segmentations highly overlapped with those of SUV4.0. LIONZ\u003csup\u003eSUV4\u003c/sup\u003e led to a stronger agreement of PET parameters and predictions with SUV4.0 compared to LIONZ. Overall, LIONZ\u003csup\u003eSUV4\u003c/sup\u003e is a suitable method for DLBCL lesion segmentation and potentially decreases reader-variability compared to threshold only based segmentation methods.\u003c/p\u003e","manuscriptTitle":"Evaluation of an artificial intelligence method for lesion segmentation of baseline FDG PET studies of DLBCL patients","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-26 09:53:03","doi":"10.21203/rs.3.rs-6294601/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b91bb1bf-a835-4b8e-85c3-a002a41824b9","owner":[],"postedDate":"March 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":46125917,"name":"Nuclear Medicine \u0026 Medical Imaging"},{"id":46125918,"name":"Artificial Intelligence and Machine Learning"},{"id":46125919,"name":"Hematology"}],"tags":[],"updatedAt":"2025-03-26T09:53:03+00:00","versionOfRecord":[],"versionCreatedAt":"2025-03-26 09:53:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6294601","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6294601","identity":"rs-6294601","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}