Edited by: Haibin Shi, Soochow University, China
Reviewed by: Hsin Wu Tseng, University of Arizona, United States; Huijie Jang, The Second Affiliated Hospital of Harbin Medical University, China
*Correspondence: Sijin Li,
†These authors have contributed equally to this work
This article was submitted to Cancer Imaging and Image-directed Interventions, a section of the journal Frontiers in Oncology
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
In this study, total lesion glycolysis (TLG) on positron emission tomography images was estimated by a trained and validated CT radiomics model, and its prognostic ability was explored among lung cancer (LC) and esophageal cancer patients (EC).
Using the identical features between the combined and thin-section CT, the estimation model of SUVsum (summed standard uptake value) was trained from the lymph nodes (LNs) of LC patients (n = 1239). Besides LNs of LC patients from other centers, the validation cohorts also included LNs and primary tumors of LC/EC from the same center. After calculating TLG (accumulated SUVsum of each individual) based on the model, the prognostic ability of the estimated and measured values was compared and analyzed.
In the training cohort, the model of 3 features was trained by the deep learning and linear regression method. It performed well in all validation cohorts (n = 5), and a linear regression could correct the bias from different scanners. Additionally, the absolute biases of the model were not significantly affected by the evaluated factors whether they included LN metastasis or not. Between the estimated natural logarithm of TLG (elnTLG) and the measured values (mlnTLG), significant difference existed among both LC (n = 137, bias = 0.510 ± 0.519, r = 0.956, P<0.001) and EC patients (n = 56, bias = 0.251± 0.463, r = 0.934, P<0.001). However, for both cancers, the overall shapes of the curves of hazard ratio (HR) against elnTLG or mlnTLG were quite alike.
Total lesion glycolysis can be estimated by three CT features with particular coefficients for different scanners, and it similar to the measured values in predicting the outcome of cancer patients.
Radiomics has gained attention recently due to the quantitative information extracted from medical images and derived features, which are associated with cancer phenotype, patient outcome, treatment response, and other classification data (
Radioactive 18F labeled Fluorodeoxyglucose (18F-FDG) is an analog of glucose, and its distribution
SUVmax of the primary tumor is commonly used in various PET/CT reports, and is related to the prognosis of myeloma, lymphoma, and other cancers (
In this study, we trained and validated a CT radiomic model to estimate the SUVsum of each lymph node or primary tumor on PET images. The prognostic ability of the estimated TLG was then evaluated and compared to the measured value.
From January 2015 to December 2019, pathologically confirmed lung cancer (LC) patients were screened from the First Affiliated Hospital of Shanxi Medical University (SMU), the First Affiliated Hospital of Anhui Medical University (AMU), and the RIDER Lung PET-CT dataset (
Eligible patients had accepted PET/CT scans (PET and combined CT), and at least one LN (short-axis diameter > 3mm) was visible on CT images identified by two experienced radiologists. For the lung cancer patients from SMU (SMU LC), besides PET/CT images, an additional thin-section chest CT scan was acquired. The excluded patients had a history of diabetes, chronic heart diseases, or chronic renal failure. The inclusion criteria of LNs were (1): in the regions (1-14) defined by the International Association for the Study of Lung Cancer (IASLC) guidelines (
The study design is presented in
Study design and variable selection strategy.
In brief, lymph nodes and primary tumors on the combined or thin-section CT images were semi-automatically segmented. The lymph nodes (LNs) of SMU LC were the training cohort. Predictable features for estimating the SUVsum of each LN were selected and sorted by deep learning method. The candidate features were the same between the top 20 features on the combined and thin-section CT images. The estimation model was trained by the partial least squares (PLS) and the linear regression of Passing & Bablok and validated using the cohorts outlined in
Training and validation cohorts.
Cohort | Scanner | Validation Aim of Model | |
---|---|---|---|
LNs of SMU LC | Training | Same | – |
LNs of SMU EC | Validation | Same | Performance in LNs of other types of cancer |
PTs of SMU LC and EC | Validation | Same | Performance in PTs |
LNs of AMU LC | Validation | Different | Performance in LNs |
LNs of RIDER LC 1 | Validation | Different | Performance in LNs |
LNs of RIDER LC 2 | Validation | Different | Performance in LNs |
LNs, lymph nodes; PTs, primary tumors; LC, lung cancer; EC, esophageal cancer; SMU, First Affiliated Hospital of Shanxi Medical University; AMU, First Hospital of Anhui Medical University; RIDER, RIDER Lung PET-CT dataset from The Cancer Imaging Archive.
For each SMU cancer patient, the total SUVsum of the primary tumor and lymph nodes was calculated by trained model, and its prognostic power was evaluated and compared with the measured TLG on PET images. The patients were followed to the end of November 2019 by medical records or communications. Overall survival (OS) was defined from the day of PET/CT scan to death for any reason. The patients were staged according to the Seventh Edition of the Union for International Cancer Control American Joint Committee on Cancer System (UICC-AJCC). Age was stratified by the median age of the cohort. Treatment modalities after PET/CT scans were stratified into no treatment, chemotherapy only, radiotherapy only, and combined chemo-radiotherapy. Target drug administration was also followed, included signal transduction inhibitors, gene expression modulators, immunotherapy drugs, and others.
Right censored survival data, and model training and validation were analyzed by the R software (version 3.6.1, a language and environment for statistical computing). Besides the basic packages of the R package, others included ggplot2, Hmisc, survcomp, h2o, and et al. Two sides of P<0.05 were considered as a significant level.
The results of model training and validation are presented in the
In the validations of the images acquired by the same scanner (
In the validations of the images acquired by different scanners, the estimation highly correlated to the corresponding measurements. However, the regression lines had slight differences that resulted from the different scanners or protocols (
Scatter plot (upper) and Bland-Altman plot (lower) in the training and validation cohorts. According to the cohorts, the scatters are grouped by colors. The x axes of the two plots are the measured ln(SUVsum). The y axes of the scatter plot and Bland-Altman plot present the estimated ln(SUVsum) and bias, respectively. The 95% CIs of regression lines between the measurements and estimations are in grey shadow, and the scatters distribution is illustrated by the box plots on the top and right sides.
Additionally, the factors of acquisition time, injected dose, volume and pathology did not significantly correlate to the estimated bias of the model (
Among the patients that could be followed, the median age of lung cancer (n=137) and esophageal cancer (n=56) patients were 64y (range: 38-88y) and 65.5y (range: 44-83y), respectively. Between the two cancers, the characteristics of pathology, TNM stage, and treatment modalities had significant differences (
Patient characteristics.
LC (n=137) | EC (n=56) | P | ||
---|---|---|---|---|
Gender | Male | 97 | 38 | 0.685 |
Female | 18 | 18 | ||
Pathology | Squamous cell carcinoma | 50 | 51 | <0.001 |
Adenocarcinoma | 76 | 5 | ||
Others | 11 | 0 | ||
Age (y) | < Median | 73 | 25 | 0.276 |
≥ Median | 64 | 31 | ||
Surgery | Yes | 41 | 40 | <0.001 |
No | 96 | 16 | ||
Radiotherapy | Yes | 47 | 46 | <0.001 |
No | 90 | 10 | ||
Chemotherapy | Yes | 88 | 46 | 0.014 |
No | 49 | 10 | ||
Target therapy | Yes | 34 | 11 | 0.440 |
No | 103 | 45 | ||
TNM stage | I-II | 34 | 26 | 0.003 |
III-IV | 103 | 30 | ||
mlnTLG | Median | 5.1 | 5.2 | 0.547 |
Range | 2.3-8.4 | 1.9-7.3 | ||
elnTLG | Median | 5.5 | 5.3 | 0.083 |
Range | 2.4-10.2 | 2.3-8.2 |
mlnTLG, Natural logarithm of measured TLG; elnTLG, Natural logarithm of estimated TLG.
To evaluate the prognostic ability of the values estimated by the model, the accumulated SUVsum of the primary tumor and lymph nodes in each patient of the SMU cohort was calculated. Between the estimated natural logarithm of TLG (elnTLG) and the measured values (mlnTLG), significant difference existed among both lung cancer (bias=0.510 ± 0.519, r=0.956, P<0.001) and esophageal cancer patients (bias=0.251± 0.463, r=0.934, P<0.001). Because the slight difference might not have a significant influence on their prognostic ability, they were analyzed as continuous variables in the following Cox regression.
After PET/CT examinations, 7 and 49 esophageal cancer patients (n=56) received surgery and combined therapy, respectively. Among the lung cancer patients, 18 did not accept any treatment, and 14, 23 and 82 individuals received surgery, chemotherapy, and combined therapy, respectively. Additionally, target drugs were administered to 34 lung cancer and 11 esophageal cancer patients. The treatment modalities had no significant difference between the pathology types of lung cancer (F=1.829, P=0.165) or esophageal cancer (F=0.137, P=0.713).
During the observation time from 30m to 106m (median: 43m), 91 lung cancer and 44 esophageal cancer patients died in the range of 0-40m (median 11 m) and 1-60m (median 15m), respectively. In the univariate Cox regression analysis (
Results of univariate Cox regression analysis for lung cancer (upper) and esophageal cancer (lower). HR with 95% CI is plotted in the fourth column.
After integrating mlnTLG or elnTLG as continuous variables, all C-index of basic models were significantly deteriorated (
C-index in 95% CI before and after integrating mlnTLG or elnTLG. The variables of basic models for lung cancer (LC) or esophageal cancer (EC) patients are separately determined by the multivariate Cox regression (in the brackets). The significance between each two C-index is indicated by *P < 0.05 or **P < 0.01.
According to the method from Liu et al. (
Plots of continuous elnTLG (upper) and mlnTLG (lower) against ln(HR) in lung cancer patients. The vertical lines and arrows indicate the median reference value of hazard ratio curves which are defined by the Akaike information criterion (AIC) values (756.11
In this study, among the lung cancer and esophageal cancer patients, the CT radiomics model trained from lymph nodes could be used to calculate the natural logarithm of SUVsum of primary tumors and lymph nodes. Furthermore, in predicting the outcome of lung cancer patients, total lesion glycolysis (TLG) calculated by the model was similar to those measured on PET images. It should be noted that, for different scanners and protocols, a simple linear regression rather than deep learning analysis could be used to correct the bias. Additionally, the model was seldom affected by the evaluated factors of injected dose, acquisition time, LN volume, pathology, and even lymph nodes metastasis or not. Above all, our results illustrated the possibility of estimating quantitative values by radiomic models.
With the development of computer science and technology, hundreds of radiomics features can be extracted from medical images. A previous study from Giesel et al. (
Recently, TLG was proposed as a better prognostic index than SUVmax among most cancer patients (
In molecular imaging-guided radiation therapy (MIGRT), the delineation of gross target volumes (GTV) on CT is improved by molecular imaging technology, such as PET and SPECT images. The simple way is to manually delineate CT GTV and revise its boundary by visual interpretation of PET (
To train a repeatable model, as pre-described in the
The advantage of the logarithmic estimation was that it was seldom affected by the evaluated factors, did and did not include lymph node metastasis. Because pathological evidence for LNs was not easy to prove, we alternately used the SUVmax=2.0 (or 2.5) as the threshold of benign and malignant ones. The correlation coefficients between the measured and estimated values were not changed according to the stratifications. Additionally, no different serial scatters appeared on the Bland-Altman and scatter plots of Model 1 (
Additionally, our data suggested that it was possible to estimate the content of lesions by a radiomics model
In summary, our study indicated that several CT radiomic features can be used to estimate the logarithm of SUVsum acquired by the same scanner, and the coefficients of the models could be corrected by a simple linear regression for different scanners or protocols.
Total lesion glycolysis can be estimated by three CT features with particular coefficients for different scanners, and it similar to the measured values in predicting the outcome of cancer patients.
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
The studies involving human participants were reviewed and approved by First Affiliated Hospital of Shanxi Medical University and the First Affiliated Hospital of Anhui Medical University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.
Conceptualization, SL. Methodology, HS and XH. Software, LZ. Validation, XX and JC. Investigation, PW and LL. Resources, ZW. Data curation, CZ. Writing—original draft preparation, HS. Writing—review and editing, SL. Supervision, S.L. Project administration, S.L. Funding acquisition, H.S. All authors contributed to the article and approved the submitted version.
The work was funded by grants from Collaborative Innovation Center for Molecular Imaging, Precise D&T Center (Grant No. MP201604) and Research Project of Health Commission of Anhui Province (Grant NO. AHWJ2021b148). Funding organizations had no role in the design, implementation, interpretation, and publication of the study.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We thank the patients who participated in this study. We would like to acknowledge The Cancer Imaging Archive.
The Supplementary Material for this article can be found online at: