Novel Functional Radiomics for Prediction of Cardiac Positron Emission Tomography Avidity in Lung Cancer Radiotherapy

Abstract

PURPOSE

Traditional methods of evaluating cardiotoxicity focus on radiation doses to the heart. Functional imaging has the potential to provide improved prediction for cardiotoxicity for patients with lung cancer. Fluorine-18 (¹⁸F) fluorodeoxyglucose (FDG)-positron emission tomography (PET)/computed tomography (CT) imaging is routinely obtained in a standard cancer staging workup. This work aimed to develop a radiomics model predicting clinical cardiac assessment using ¹⁸F-FDG PET/CT scans before thoracic radiation therapy.

METHODS

Pretreatment ¹⁸F-FDG PET/CT scans from three study populations (N = 100, N = 39, N = 70) were used, comprising two single-institutional protocols and one publicly available data set. A clinician (V.J.) classified the PET/CT scans per clinical cardiac guidelines as no uptake, diffuse uptake, or focal uptake. The heart was delineated, and 210 novel functional radiomics features were selected to classify cardiac FDG uptake patterns. Training data were divided into training (80%)/validation (20%) sets. Feature reduction was performed using the Wilcoxon test, hierarchical clustering, and recursive feature elimination. Ten-fold cross-validation was carried out for training, and the accuracy of the models to predict clinical cardiac assessment was reported.

RESULTS

From 202 of 209 scans, cardiac FDG uptake was scored as no uptake (39.6%), diffuse uptake (25.3%), and focal uptake (35.1%), respectively. Sixty-two independent radiomics features were reduced to nine clinically pertinent features. The best model showed 93% predictive accuracy in the training data set and 80% and 92% predictive accuracy in two external validation data sets.

CONCLUSION

This work used an extensive patient data set to develop a functional cardiac radiomic model from standard-of-care ¹⁸F-FDG PET/CT scans, showing good predictive accuracy. The radiomics model has the potential to provide an automated method to predict existing cardiac conditions and provide an early functional biomarker to identify patients at risk of developing cardiac complications after radiotherapy.

Novel radiomics model predicts cardiac toxicity from 18F-FDG PET/CT scans with high accuracy.

INTRODUCTION

Lung cancer is a leading cause of cancer-related deaths worldwide,¹ and definitive (chemo-)radiotherapy is a standard treatment option for patients with locally advanced disease. However, radiotherapy can also cause cardiac toxicity, associated with an increased risk of mortality and morbidity.² Studies have highlighted the clinical impact of the radiation dose to the heart for patients with lung cancer receiving radiation therapy.^3,4 There has been a noted link between cardiac dose and overall survival (OS),^4-6 cardiotoxicity,⁷ and cardiac mortality.⁸ Patients who receive higher radiation doses to the heart are likely to have decreased OS,⁴ increased probability of cardiac events,^7,9 and increased rates of death due to cardiac events.⁸ Therefore, for patients with lung cancer receiving chemoradiation, it is clinically important to identify patients with baseline cardiac disease or suboptimal cardiac function and to be able to provide an early prediction for which patients may be at an increased risk of cardiac toxicity.

CONTEXT

• Key Objective

• This study pioneers the use of functional radiomics for predicting cardiac positron emission tomography (PET) avidity in patients with lung cancer undergoing radiotherapy.

• Knowledge Generated

• Our study introduces a novel radiomics model using fluorine-18 (¹⁸F) fluorodeoxyglucose (FDG) PET/computed tomography (CT) scans, demonstrating a robust 93% predictive accuracy in training data and 80%-92% accuracy in external validation sets. This innovative approach provides an automated method to identify existing cardiac conditions and serves as an early functional biomarker for assessing the risk of cardiac complications postradiotherapy.

• Relevance

• Using ¹⁸F-FDG PET/CT scans before thoracic radiation therapy, this work created a radiomics model for evaluating the heart. The model could offer a way to automatically detect current heart problems and provide an early functional biomarker to identify patients at risk of developing cardiac complications after radiotherapy.

Traditional radiotherapy methods of evaluating, mitigating, and assessing for cardiotoxicity have focused on dosimetric evaluation of radiation doses to the heart (eg, mean heart dose).⁶ A critical shortcoming of these prediction models is that they determine population-based cardiotoxicity risk and often need to provide accurate and robust risk assessment for most patients.^6,10 A method with the potential for improving the prediction of cardiotoxicity centers on the idea of incorporating functional imaging into the assessment paradigm. Studies across multiple disease sites and clinical end points have underscored how functional imaging can significantly improve the prediction of both the OS¹¹ and toxicity.^12,13 An imaging modality with the potential to provide clinically pertinent cardiac information is fluorine-18 (¹⁸F) fluorodeoxyglucose (FDG) positron emission tomography (PET) imaging. Although FDG PET imaging is primarily thought of as a way to image the tumor,¹⁴ it is also a standard imaging modality used to assess the heart and a clinically accepted way to image for cardiac inflammation and myocardial viability.¹⁵ Recent studies have used standard-of-care staging and disease assessment FDG PET-computed tomography (PET/CT) scans to show that changes in the standard uptake value (SUV) in the heart from pretreatmet to post-treatment are predictive of clinical outcomes.^16-18 Clinical guidelines¹⁵ note that myocardial FDG uptake can predict myocardial viability, myocarditis, cardiac sarcoidosis, and cardiovascular infections. Standard-of-care FDG PET/CT scans have the potential to provide clinically pertinent cardiac information without adding an extra imaging procedure for the patient. However, cardiac assessment of oncologic staging and follow-up FDG PET/CT scans is not routinely performed. Automated techniques to classify myocardial uptake patterns on the basis of clinical guidelines can provide seminal clinical cardiac data by repurposing standard staging and follow-up oncologic PET/CT scans.

Radiomics is an emerging field that involves extracting and analyzing quantitative features from medical images to predict clinical outcomes.^19-21 Recently, radiomics models have been developed to predict treatment response and prognosis in various cancers, including lung cancer.^22-34 The potential of radiomics models to evaluate the cardiac SUV signal has yet to be evaluated. Therefore, this study's purpose was to use a 209-patient data set from three sources to develop and evaluate a radiomics model to interpret cardiac FDG uptake patterns on pretreatment PET images in patients with lung cancer and predict cardiac clinical assessment prior to thoracic radiotherapy.

METHODS

The study included 209 pretreatment ¹⁸F-FDG PET-CT images of patients with lung cancer from three different study populations. Patients with lung cancer treated with radiotherapy at Thomas Jefferson University (TJU) in Philadelphia, PA, from October 2020 to May 2022 underwent a systematic review. One hundred eligible cases were chosen, meeting specific inclusion criteria that mandated access to pretreatment ¹⁸F-FDG PET/CT. The data set derived from TJU was subsequently used as the training data set. The institutional review board approved the study (21F.644). Two external validation data sets were used: American College of Radiology Imaging Network (ACRIN) 6668³⁵ (N = 70)³⁶ and the University of Colorado (CU) data (Colorado Multiple Institutional Review Board [COMIRB] 14-1856; N = 39) previously published.¹⁶ Patients with poor PET/CT image quality (reviewed by study authors) were excluded from this work (details presented in the Clinical Cardiac Assessment of PET Scans section).

The proposed study presents a predictive model that can predict cardiac uptake on FDG-PET using radiomics. This study consists of a clinical cardiac assessment of PET scans by a clinician (V.J.), radiomics feature extraction, feature selection, model building, and evaluation, as shown in Figure 1. The details are discussed in the following subsections.

FIG 1.

Study and cardiac model diagram, training data set: TJU (N = 100), and external validation data sets: ACRIN (N = 70) and CU (N = 39). ACRIN, American College of Radiology Imaging Network; CU, University of Colorado; CV, cross-validation; FDG, fluorodeoxyglucose; LV, left ventricular; RFE, recursive feature elimination; SFS, sequential feature selection; TJU, Thomas Jefferson University; TPOT, tree-based optimization tool for automating machine learning.

Clinical Cardiac Assessment of PET Scans

On the basis of a clinical scoring system of the FDG uptake in the heart presented by clinical guidelines¹⁵ and previous studies of FDG imaging of cardiac conditions,³⁷ a clinician reviewed and scored the PET scans. First, the clinician reviewed whether the PET image had inadequate quality or excessive artifacts. Seven scans were deemed to have insufficient image quality (TJU: 1, ACRIN: 4, CU: 2). Once a PET scan was considered to have sufficient quality, the clinician reviewed the cardiac portion of the PET scan with the following scoring criteria on the basis of previously published methods^15,37: no myocardial uptake, diffuse myocardial uptake, and focal myocardial uptake as shown in Figure 2.

FIG 2.

Cardiac uptake classifications: no uptake, diffuse uptake, and focal uptake.

Radiomics Feature Extraction

We used cardiac FDG-PET/CT radiomics to analyze FDG uptake patterns of the heart. The heart was delineated on the CT from the FDG-PET/CT by an experienced medical physicist following the RTOG heart atlas. In addition, to evaluate feature robustness and a completely automated solution, an AI autocontouring tool³⁸ was used to segment the heart. Once the heart was contoured on the CT from the FDG-PET/CT, the contours were transferred to the corresponding PET images. The heart contours were visually reviewed on both the CT and PET images to confirm there was no misalignment of the images. We extracted 210 radiomics features from the PET image within the heart contour. PyRadiomics software (version 3.0.1)³⁹ and in-house software^22,40 were used to extract the features, which included first-order statistics, texture features such as gray level co-occurrence matrix (GLCM), gray level size zone matrix, gray level run length matrix, neighboring gray-tone difference matrix, gray level dependence matrix (GLDM), and 2D and 3D shape features. The Data Supplement (Appendix SA) thoroughly explains the radiomics features. Feature extraction using PyRadiomics is detailed in the Data Supplement (Appendix SA1), and the Data Supplement (Appendix SA2) provides details about our in-house radiomics tool-based feature extraction.

Feature Selection

We divided training data (TJU) into an 80% training set and a 20% validation set to train and validate the classification models. Before building classification models, we used the intraclass correlation coefficient (ICC) between features extracted by manual and automated segmentation to filter out nonrobust features when the ICC is <0.9. For the remaining features, we assessed the predictive capability of each feature using the Wilcoxon test and identified good predictors on the basis of a Bonferroni-adjusted P value of <.05. We removed highly correlated features through hierarchical clustering to minimize collinearity. The remaining features were used in the model-building process. We performed further feature selection in the prediction pipeline, discussed in the next section.

Model Building and Evaluation

We employed the tree-based pipeline optimization tool for automating machine learning (TPOT) AutoML framework⁴¹ to discover preliminary classification models. TPOT is a Python library that automates the machine learning pipeline and uses genetic programming, an optimization technique, to build the best program by mimicking the natural selection process of evolution. We selected the best model on the basis of the highest accuracy score on the validation set through the three-step prediction pipeline optimization using TPOT framework, which are (1) model discovery, (2) hyperparameter optimization, and (3) feature optimization. The model-building steps are clearly outlined in the Data Supplement (Appendix SB). Specifically, the Data Supplement (Appendix SB1) outlines the model building process, (Appendix SB2) provides a comprehensive overview of the available methods to construct pipelines, and (Appendix SB3) includes detailed information on model evaluation. We evaluated the model accuracy on the training data (TJU) and on two external validation data sets (ACRIN and CU).

RESULTS

Cardiac FDG Uptake Pattern Classification

TJU, ACRIN, and CU data sets have 100, 70, and 39 cases (total 209 cases), respectively. Owing to poor image quality, one case was excluded from the TJU data set, four from ACRIN, and two from CU. Therefore, a total of 202 cases were enrolled. Table 1 presents the distribution of the clinical grading for the myocardial SUV uptake for each data set. Among them, 80 cases (39.6%) were scored as no uptake, 51 cases (25.3%) were scored as diffuse uptake, and 71 cases (35.1%) were scored as focal uptake.

TABLE 1.

Statistics of LV FDG Uptake Pattern Classification

Class	LV Myocardial Uptake Pattern	TJU, No. (%)	ACRIN, No. (%)	CU, No. (%)	Total, No. (%)
0	No uptake	44 (44.4)	23 (34.8)	13 (35.1)	80 (39.6)
1	Diffuse uptake	19 (19.2)	21 (31.8)	11 (29.8)	51 (25.3)
2	Focal uptake	36 (36.4)	22 (33.3)	13 (35.1)	71 (35.1)
All		99	66	37	202

Abbreviations: ACRIN, American College of Radiology Imaging Network; CU, University of Colorado; FDG, fluorodeoxyglucose; LV, left ventricular; TJU, Thomas Jefferson University.

Radiomics Feature Selection and Model Building

A total of 200 radiomic features were extracted from PET images. One hundred and fifty-three robust features were selected after ICC filtering, and 141 features were retained after Wilcoxon test with Bonferroni correction. Hierarchical clustering resulted in selecting 62 independent features to avoid collinearity. Table 2 shows the iteration and optimization of the radiomics model pipeline with accompanying accuracy results.

TABLE 2.

Prediction Model Pipeline Optimization Results and Validation Results

Model Building Steps	No. of Features	Accuracy, %
		10-Fold CV	Training	Validation	External Validation ACRIN	External Validation CU
Iteration 1: model discovery	56	89.9	94.9	90.0	78.8	89.2
Iteration 2: hyper parameter optimization	52	89.9	96.2	90.0	78.8	91.9
Iteration 3: feature optimization	9	89.8	92.4	95.0	80.3	91.9

NOTE. Bold values represent the maximum accuracy achieved for each evaluation type across the iterations.

Abbreviations: ACRIN, American College of Radiology Imaging Network; CU, University of Colorado; CV, cross-validation.

The model pipeline discovery was applied upon the 62 features. The TPOT discovered the best prediction model pipeline, Variance Threshold-One Hot Encoder-Extra Trees Classifier, as shown in Figure 3. The result of the 10-fold cross-validation was an average accuracy of 88.9%. The variance threshold selected 56 features. The training set had an accuracy of 94.9%, and the validation set had an accuracy of 90%. Two external validations showed accuracies of 78.8% and 89.2%.

FIG 3.

Optimization steps of the prediction model pipeline with selected methods for each component with optimized hyperparameters and their validations. RFE, recursive feature elimination; SFS, sequential feature selection; TJU, Thomas Jefferson University; TPOT, tree-based optimization tool for automating machine learning.

After hyperparameter optimization, the pipeline is updated as shown in the second row of Figure 3. The best 10-fold cross-validation result was an average accuracy of 89.9%. The selected features were reduced from 56 to 52 features, and the training set performance was improved to 96.2%. The validation accuracy remained the same at 90%, and the external validation accuracy was 78.8% and 91.9% (CU).

On the basis of the discovered prediction pipeline, we optimized the number of features. Nine features were selected as the best that maintained performance while reducing the number of features. The optimized pipeline is shown in the last row in Figure 3. The best 10-fold cross-validation result was an average accuracy of 89.8%. The training set showed an accuracy of 92.4%; the validation set 95.0%, the external validation on ACRIN 80.3%, and CU 91.9%. The nine features were PET 2D Maximum, PET 2D Skewness, PET Mean, PET GLCM Cluster Prominence, PET GLCM Sum Average, PET GLCM Cluster Shade, PET GLDM Dependence Variance, PET Maximum, and PET 2D Kurtosis in order of importance ranking.

The performance evaluation of the optimized model is available in the Data Supplement (Appendix SB3). The Data Supplement (Figure SB1) presents accuracy, precision, recall, and f1-scores by class, along with their macroaverage and weighted average, for the optimized model across data sets TJU, ACRIN, and CU. Furthermore, the Data Supplement (Appendix SC) presents the results for automated heart contouring, encompassing feature selection, model construction, and performance evaluation of the optimized model. The results obtained through automated contouring are comparable with those obtained through manual contouring, as demonstrated in the Data Supplement (Table SC1 and Figures SC2 and SC3).

DISCUSSION

This work aimed to develop and evaluate a novel radiomics model to predict cardiac FDG uptake patterns using standard-of-care pretreatment PET/CT staging scans for patients with lung cancer. The study found that while a quarter of the patients had no FDG uptake in the heart, about half had homogenous FDG uptake, and a quarter had heterogeneous FDG uptake. On the basis of clinical guidelines,^15,37 both homogenous and heterogeneous FDG uptake could be considered abnormal. Studies have shown that FDG uptake is a marker of cardiac inflammation, and specific uptake patterns can be predictive of cardiac abnormalities including heterogeneous uptake demonstrating radiation-induced pericarditis, ventricular myocardium uptake predicting for cardiomyopathy, and focal uptake demonstrating sarcoidosis.⁴² Our data showed that cardiac radiomics features, using either manual or automated heart contours, can predict clinical interpretation of the cardiac portion of FDG PET/CT scans with over 90% accuracy when validated in the TJU data set. The external validation of the CU data set showed an accuracy of about 90% when both heart contours were used. However, the external validation of the ACRIN data set showed relatively low performance. This may be attributed to the data set being collected approximately 10 years ago, resulting in poorer image quality compared with other data sets. Additionally, twice as many images (n = 4) were excluded because of poor image quality. Clinical guidelines^15,37 and previous studies have shown that FDG PET/CT scans can provide clinically pertinent cardiac information.^16-18 Our work provides the first use of radiomics to generate an automated method to evaluate the cardiac aspect of standard-of-care FDG PET/CT scans. The novelty of this work is that a large patient data set was used to develop a robust, automated radiomics model to predict cardiac clinical interpretations of FDG PET scans with reasonable accuracy as validated on independent cohorts. If further validated, this radiomics model, when combined with another patient, clinical, and treatment parameters, can provide an automated method to use the standard-of-care staging FDG PET/CT scans to predict both existing cardiac conditions and provide an early functional biomarker to identify patients at risk of developing cardiac complications after radiotherapy.

Previous work has explored using standard-of-care FDG PET/CT scans to predict normal tissue toxicities. For patients with head and neck cancer, van Dijk et al¹³ demonstrated that pretreatment parotid SUV features significantly improved the prediction of xerostomia. Similarly, studies in patients with lung cancer have shown that changes in lung SUV as a function of radiotherapy doses^41-43 and pretreatment SUV lung features⁴³ were predictive of symptomatic radiation pneumonitis. In the heart, early studies characterized a dose-response^44,45 as a function of radiation, and subsequent work showed that uptake changes in the heart from pretreatment to post-treatment are predictive for OS for both patients with lung cancer^16,18 and patients with esophageal cancer.¹⁷ With regards to quantitative data, studies showed that for patients with lung cancer receiving thoracic radiotherapy, there is an average increase of SUV of 1.7% in the heart per 10 Gy and that patients who were alive at last follow-up had an increase of 17.2% in cardiac SUV compared with a decrease of 13.5% in cardiac SUV for patients who were not alive at last follow-up.¹⁶ This study builds on previous quantitative analysis by extending the work to features beyond average SUV in the heart. This work aligns with previous studies in that it characterizes clinically meaningful information from the PET signal in the heart and extends previous literature by demonstrating how radiomics can be applied to cardiac functional imaging.

This study is retrospective and has several limitations related to the retrospective nature of the work. The FDG PET/CT scans used for this work were acquired for cancer staging and were not purposefully collected, limiting the uniformity of the data. Although the contrast and imaging protocols for oncologic staging FDG PET/CT scans are similar to cardiac PET imaging, the fasting guidelines differ for the two types of scans (patients undergoing cardiac-focused FDG PET/CT scans are directed to go on a lower carbohydrate diet while no such dietary changes are recommended to oncologic patients). Future work will evaluate and characterize the differences in the cardiac signal between cardiac-focused FDG PET/CT scans and oncologic FDG PET/CT scans. Since the PET/CT scans were not prospectively acquired, the imaging data for this work came from different PET scanners. Previous studies have noted that absolute SUV can vary between scanners.⁴⁶ Additionally, information about preexisting clinical cardiac conditions and comorbidities was unavailable. Subsequent work will collect clinical cardiac information and evaluate whether the radiomics model can predict preexisting cardiac conditions and combine with other treatment and patient-related factors (radiation dose, concomitant/neoadjuvant/adjuvant therapies, oncologic stage, baseline pulmonary, and cardiac comorbidities) to predict for post-treatment cardiac complications. This study is the first to present the concept of functional cardiac radiomics and provide seminal data for hypothesis-generating research using an automated cardiac uptake functional biomarker to predict clinical outcomes.

In the development of our radiomics model, we used a manual contour of the heart. Additionally, we incorporated an autocontouring software named INTContour for comparison. There are numerous autocontouring software options available, ranging from commercial solutions like Contour ProtégéAI (MIM Software⁴⁷) to open-source solutions.⁴⁸ Our findings revealed that both manual and autocontouring methods yielded comparable performance with a similar set of features selected. This suggests that the process can be replicated with any robust heart contour AI model.

For the early detection of cardiac toxicity, dual-energy CT (DECT)⁴⁹ and magnetic resonance imaging (MRI)⁵⁰ serve as alternative radiomic approaches. DECT is particularly useful for its material differentiation capabilities, which include myocardial perfusion and iodine mapping, because of its dual-energy imaging technique. Conversely, MRI is renowned for its superior soft tissue contrast and does not involve ionizing radiation, making it a comprehensive tool for evaluating cardiac function, including aspects such as myocardial perfusion, viability, and wall motion. While both DECT and MRI have shown promise in the early prediction of cardiac toxicity, it is crucial to note that they are not yet standard imaging modalities in lung cancer radiotherapy protocols. Despite their potential advantages, their role in standard radiotherapy remains to be fully established.

This study used a large patient data set to develop and evaluate a functional cardiac radiomics model to predict for clinical interpretation of standard-of-care FDG PET/CT scans. The data demonstrated the best model with the nine optimal features that could predict the clinical characterization of the cardiac aspect of the FDG PET/CT scans with an accuracy of 93% on TJU. External validation on two independent data sets resulted in an accuracy of 80% (ACRIN) and 92% (CU). The radiomics model can provide an automated method to use the standard of care staging FDG PET/CT scans to predict both existing cardiac conditions and provide an early functional biomarker to identify patients at risk of developing cardiac complications after radiotherapy.

AUTHOR CONTRIBUTIONS

Conception and design: Wookjin Choi, Adam P. Dicker, Nicole L. Simone, Yevgeniy Vinogradskiy

Administrative support: Adam P. Dicker

Provision of study materials or patients: Yingcui Jia, Maria Werner-Wasik

Collection and assembly of data: Wookjin Choi, Yingcui Jia, Jennifer Kwak, Maria Werner-Wasik, Varsha Jain, Yevgeniy Vinogradskiy

Data analysis and interpretation: Wookjin Choi, Adam P. Dicker, Nicole L. Simone, Eugene Storozynsky, Varsha Jain, Yevgeniy Vinogradskiy

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).