Song Chen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Song Chen has a total of 64 co-authors. Here are the top co-authors by number of shared publications:
- Timothy Perk
AIQ Solutions, Madison, WI 53717, USA
4 shared publications - Yaming Li
Department of Nuclear Medicine, The First Hospital of China Medical University, Shenyang, People’s Republic of China
4 shared publications Robert Jeraj
Department of Medical Physics, Wisconsin Institutes for Medical Research, University of Wisconsin-Madison, Madison, USA
3 shared publications- Xuena Li
Department of Nuclear Medicine, the First Hospital of China Medical University, Shenyang, China
3 shared publications Bosinan Chen
The First Hospital of China Medical University, Shenyang, China
2 shared publicationsGlenn Liu
University of Wisconsin Carbone Cancer Center, University of Wisconsin Madison, Madison, UNITED STATES
2 shared publications
Latest Publications
- Open Access
- 31 August 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Clinical attribute-based risk matrix for prognostication in immunotherapy-treated non-small cell lung cancer
- 31 August 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessClinical attribute-based risk matrix for prognostication in immunotherapy-treated non-small cell lung cancer
- Open Access
- 20 December 2025
- Published by Springer Nature in npj Digital Medicine
- Vol. 9 (1)
Predicting Invasiveness of Lung Adenocarcinoma from Chest CT with Few-shot Vision-Language Ternary Classification Model
Preoperative differentiation of preinvasive lesions, minimally invasive adenocarcinomas, and invasive adenocarcinomas within pure ground-glass nodules (pGGNs) is challenging. Herein, this study investigated the potential of vision-language models to assist radiologists in noninvasively predicting pGGN invasiveness on CT scans. This retrospective multicenter study enrolled 848 patients with pathologically-confirmed lung adenocarcinoma manifesting as pGGNs. GPT-4o was tasked with localizing pGGNs on CT scans to detect ten pGGN invasiveness-associated features to generate a diagnosis and was compared with Molmo. The twenty-shot GPT-4o model demonstrated superior performance in the ternary classification of pGGN invasiveness (Delong test, P < 0.01). Six radiologists’ assessments revealed that GPT-4o output showed high reliability, willingness to use, reliance, low risk of harm, inappropriate content, and missing content. With GPT-4o assistance, another six radiologists achieved an average improvement in pGGN invasiveness diagnosis. The twenty-shot-based GPT-4o model exhibited superior diagnostic capability for pGGN invasiveness in lung adenocarcinoma, achieving significantly improved diagnostic accuracy by radiologists.
- 20 December 2025
- Published by Springer Nature in npj Digital Medicine
- Vol. 9 (1)
Open AccessPredicting Invasiveness of Lung Adenocarcinoma from Chest CT with Few-shot Vision-Language Ternary Classification Model
Preoperative differentiation of preinvasive lesions, minimally invasive adenocarcinomas, and invasive adenocarcinomas within pure ground-glass nodules (pGGNs) is challenging. Herein, this study investigated the potential of vision-language models to assist radiologists in noninvasively predicting pGGN invasiveness on CT scans. This retrospective multicenter study enrolled 848 patients with pathologically-confirmed lung adenocarcinoma manifesting as pGGNs. GPT-4o was tasked with localizing pGGNs on CT scans to detect ten pGGN invasiveness-associated features to generate a diagnosis and was compared with Molmo. The twenty-shot GPT-4o model demonstrated superior performance in the ternary classification of pGGN invasiveness (Delong test, P < 0.01). Six radiologists’ assessments revealed that GPT-4o output showed high reliability, willingness to use, reliance, low risk of harm, inappropriate content, and missing content. With GPT-4o assistance, another six radiologists achieved an average improvement in pGGN invasiveness diagnosis. The twenty-shot-based GPT-4o model exhibited superior diagnostic capability for pGGN invasiveness in lung adenocarcinoma, achieving significantly improved diagnostic accuracy by radiologists.
- Open Access
- 1 May 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Imaging Cancer
- Vol. 7 (3) , e240253
Evaluating Automated Tools for Lesion Detection on 18F Fluoroestradiol PET/CT Images and Assessment of Concordance with Standard-of-Care Imaging in Metastatic Breast Cancer
An artificial intelligence model was trained to detect lesions on fluorine 18 (18F)-fluoroestradiol (FES) PET/CT images and an automated concordance tool measured heterogeneity between 18F-FES PET/CT and standard-of-care imaging.
- 1 May 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Imaging Cancer
- Vol. 7 (3) , e240253
Open AccessEvaluating Automated Tools for Lesion Detection on 18F Fluoroestradiol PET/CT Images and Assessment of Concordance with Standard-of-Care Imaging in Metastatic Breast Cancer
An artificial intelligence model was trained to detect lesions on fluorine 18 (18F)-fluoroestradiol (FES) PET/CT images and an automated concordance tool measured heterogeneity between 18F-FES PET/CT and standard-of-care imaging.
- 18 October 2023
- Published by IOP Publishing in Biomedical Physics & Engineering Express
- Vol. 9 (6) , 065021
Multi-organ segmentation of CT via convolutional neural network: impact of training setting and scanner manufacturer
Objective. Automated organ segmentation on CT images can enable the clinical use of advanced quantitative software devices, but model performance sensitivities must be understood before widespread adoption can occur. The goal of this study was to investigate performance differences between Convolutional Neural Networks (CNNs) trained to segment one (single-class) versus multiple (multi-class) organs, and between CNNs trained on scans from a single manufacturer versus multiple manufacturers. Methods. The multi-class CNN was trained on CT images obtained from 455 whole-body PET/CT scans (413 for training, 42 for testing) taken with Siemens, GE, and Phillips PET/CT scanners where 16 organs were segmented. The multi-class CNN was compared to 16 smaller single-class CNNs trained using the same data, but with segmentations of only one organ per model. In addition, CNNs trained on Siemens-only (N = 186) and GE-only (N = 219) scans (manufacturer-specific) were compared with CNNs trained on data from both Siemens and GE scanners (manufacturer-mixed). Segmentation performance was quantified using five performance metrics, including the Dice Similarity Coefficient (DSC). Results. The multi-class CNN performed well compared to previous studies, even in organs usually considered difficult auto-segmentation targets (e.g., pancreas, bowel). Segmentations from the multi-class CNN were significantly superior to those from smaller single-class CNNs in most organs, and the 16 single-class models took, on average, six times longer to segment all 16 organs compared to the single multi-class model. The manufacturer-mixed approach achieved minimally higher performance over the manufacturer-specific approach. Significance. A CNN trained on contours of multiple organs and CT data from multiple manufacturers yielded high-quality segmentations. Such a model is an essential enabler of image processing in a software device that quantifies and analyzes such data to determine a patient’s treatment response. To date, this activity of whole organ segmentation has not been adopted due to the intense manual workload and time required.
- 18 October 2023
- Published by IOP Publishing in Biomedical Physics & Engineering Express
- Vol. 9 (6) , 065021
Multi-organ segmentation of CT via convolutional neural network: impact of training setting and scanner manufacturer
Objective. Automated organ segmentation on CT images can enable the clinical use of advanced quantitative software devices, but model performance sensitivities must be understood before widespread adoption can occur. The goal of this study was to investigate performance differences between Convolutional Neural Networks (CNNs) trained to segment one (single-class) versus multiple (multi-class) organs, and between CNNs trained on scans from a single manufacturer versus multiple manufacturers. Methods. The multi-class CNN was trained on CT images obtained from 455 whole-body PET/CT scans (413 for training, 42 for testing) taken with Siemens, GE, and Phillips PET/CT scanners where 16 organs were segmented. The multi-class CNN was compared to 16 smaller single-class CNNs trained using the same data, but with segmentations of only one organ per model. In addition, CNNs trained on Siemens-only (N = 186) and GE-only (N = 219) scans (manufacturer-specific) were compared with CNNs trained on data from both Siemens and GE scanners (manufacturer-mixed). Segmentation performance was quantified using five performance metrics, including the Dice Similarity Coefficient (DSC). Results. The multi-class CNN performed well compared to previous studies, even in organs usually considered difficult auto-segmentation targets (e.g., pancreas, bowel). Segmentations from the multi-class CNN were significantly superior to those from smaller single-class CNNs in most organs, and the 16 single-class models took, on average, six times longer to segment all 16 organs compared to the single multi-class model. The manufacturer-mixed approach achieved minimally higher performance over the manufacturer-specific approach. Significance. A CNN trained on contours of multiple organs and CT data from multiple manufacturers yielded high-quality segmentations. Such a model is an essential enabler of image processing in a software device that quantifies and analyzes such data to determine a patient’s treatment response. To date, this activity of whole organ segmentation has not been adopted due to the intense manual workload and time required.
- 1 November 2022
- Published by Springer Nature in Chinese Journal of Academic Radiology
- Vol. 5 (4) , 238-243
Using gray level co-occurrence matrix texture features extracted from dual time point PET/CT images to diagnose malignant FDG-avid solitary pulmonary nodules
- 1 November 2022
- Published by Springer Nature in Chinese Journal of Academic Radiology
- Vol. 5 (4) , 238-243
Using gray level co-occurrence matrix texture features extracted from dual time point PET/CT images to diagnose malignant FDG-avid solitary pulmonary nodules
- Open Access
- 16 August 2019
- Published by Springer Nature in Cancer Imaging
- Vol. 19 (1) , 1-8
Using neighborhood gray tone difference matrix texture features on dual time point PET/CT images to differentiate malignant from benign FDG-avid solitary pulmonary nodules
Lung cancer usually presents as a solitary pulmonary nodule (SPN) on diagnostic imaging during the early stages of the disease. Since the early diagnosis of lung cancer is very important for treatment, the accurate diagnosis of SPNs has much importance. The aim of this study was to evaluate the discriminant power of dual time point imaging (DTPI) PET/CT in the differentiation of malignant and benign FDG-avid solitary pulmonary nodules by using neighborhood gray-tone difference matrix (NGTDM) texture features. Retrospective analysis was carried out on 116 patients with SPNs (35 benign and 81 malignant) who had DTPI 18F-FDG PET/CT between January 2005 and May 2015. Both PET and CT images were acquired at 1 h and 3 h after injection. The SUVmax and NGTDM texture features (coarseness, contrast, and busyness) of each nodule were calculated on dual time point images. Patients were randomly divided into training and validation datasets. Receiver operating characteristic (ROC) curve analysis was performed on all texture features in the training dataset to calculate the optimal threshold for differentiating malignant SPNs from benign SPNs. For all the lesions in the testing dataset, two visual interpretation scores were determined by two nuclear medicine physicians based on the PET/CT images with and without reference to the texture features. In the training dataset, the AUCs of delayed busyness, delayed coarseness, early busyness, and early SUVmax were 0.87, 0.85, 0.75 and 0.75, respectively. In the validation dataset, the AUCs of visual interpretations with and without texture features were 0.89 and 0.80, respectively. Compared to SUVmax or visual interpretation, NGTDM texture features derived from DTPI PET/CT images can be used as good predictors of SPN malignancy. Improvement in discriminating benign from malignant nodules using SUVmax and visual interpretation can be achieved by adding busyness extracted from delayed PET/CT images.
- 16 August 2019
- Published by Springer Nature in Cancer Imaging
- Vol. 19 (1) , 1-8
Open AccessUsing neighborhood gray tone difference matrix texture features on dual time point PET/CT images to differentiate malignant from benign FDG-avid solitary pulmonary nodules
Lung cancer usually presents as a solitary pulmonary nodule (SPN) on diagnostic imaging during the early stages of the disease. Since the early diagnosis of lung cancer is very important for treatment, the accurate diagnosis of SPNs has much importance. The aim of this study was to evaluate the discriminant power of dual time point imaging (DTPI) PET/CT in the differentiation of malignant and benign FDG-avid solitary pulmonary nodules by using neighborhood gray-tone difference matrix (NGTDM) texture features. Retrospective analysis was carried out on 116 patients with SPNs (35 benign and 81 malignant) who had DTPI 18F-FDG PET/CT between January 2005 and May 2015. Both PET and CT images were acquired at 1 h and 3 h after injection. The SUVmax and NGTDM texture features (coarseness, contrast, and busyness) of each nodule were calculated on dual time point images. Patients were randomly divided into training and validation datasets. Receiver operating characteristic (ROC) curve analysis was performed on all texture features in the training dataset to calculate the optimal threshold for differentiating malignant SPNs from benign SPNs. For all the lesions in the testing dataset, two visual interpretation scores were determined by two nuclear medicine physicians based on the PET/CT images with and without reference to the texture features. In the training dataset, the AUCs of delayed busyness, delayed coarseness, early busyness, and early SUVmax were 0.87, 0.85, 0.75 and 0.75, respectively. In the validation dataset, the AUCs of visual interpretations with and without texture features were 0.89 and 0.80, respectively. Compared to SUVmax or visual interpretation, NGTDM texture features derived from DTPI PET/CT images can be used as good predictors of SPN malignancy. Improvement in discriminating benign from malignant nodules using SUVmax and visual interpretation can be achieved by adding busyness extracted from delayed PET/CT images.
- 1 November 2018
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 63 (22) , 225019
Automated classification of benign and malignant lesions in 18 F-NaF PET/CT images using machine learning
Purpose . 18 F-NaF PET/CT imaging of bone metastases is confounded by tracer uptake in benign diseases, such as osteoarthritis. The goal of this work was to develop an automated bone lesion classification algorithm to classify lesions in NaF PET/CT images. Methods . A nuclear medicine physician manually identified and classified 1751 bone lesions in NaF PET/CT images from 37 subjects with metastatic castrate-resistant prostate cancer, 14 of which (598 lesions) were analyzed by three additional physicians. Lesions were classified on a five-point scale from definite benign to definite metastatic lesions. Classification agreement between physicians was assessed using Fleiss’ κ . To perform fully automated lesion classification, three different lesion detection methods based on thresholding were assessed: SUV > 10 g ml −1 , SUV > 15 g ml −1 , and a statistically optimized regional thresholding (SORT) algorithm. For each ROI in the image, 172 different imaging features were extracted, including PET, CT, and spatial probability features. These imaging features were used as inputs into different machine learning algorithms. The impact of different deterministic factors affecting classification performance was assessed. Results . The factors that most impacted classification performance were the machine learning algorithm and the lesion identification method. Random forests (RF) had the highest classification performance. For lesion segmentation, using SORT (AUC = 0.95 [95%CI = 0.94–0.95], sensitivity = 88% [86%–90%], and specificity = 0.89 [0.87–0.90]) resulted in superior classification performance ( p < 0.001) compared to SUV > 10 g ml −1 (AUC = 0.87) and SUV > 15 g ml −1 (AUC = 0.86). While there was only moderate agreement between physicians in lesion classification ( κ = 0.53 [95% CI = 0.52–0.53]), classification performance was high using any of the four physicians as ground truth (AUC range: 0.91–0.93). Conclusion . We have developed the first whole-body automatic disease classification tool for NaF PET using RF, and demonstrated its ability to replicate different physicians’ classification tendencies. This enables fully-automated analysis of whole-body NaF PET/CT images.
- 1 November 2018
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 63 (22) , 225019
Automated classification of benign and malignant lesions in 18 F-NaF PET/CT images using machine learning
Purpose . 18 F-NaF PET/CT imaging of bone metastases is confounded by tracer uptake in benign diseases, such as osteoarthritis. The goal of this work was to develop an automated bone lesion classification algorithm to classify lesions in NaF PET/CT images. Methods . A nuclear medicine physician manually identified and classified 1751 bone lesions in NaF PET/CT images from 37 subjects with metastatic castrate-resistant prostate cancer, 14 of which (598 lesions) were analyzed by three additional physicians. Lesions were classified on a five-point scale from definite benign to definite metastatic lesions. Classification agreement between physicians was assessed using Fleiss’ κ . To perform fully automated lesion classification, three different lesion detection methods based on thresholding were assessed: SUV > 10 g ml −1 , SUV > 15 g ml −1 , and a statistically optimized regional thresholding (SORT) algorithm. For each ROI in the image, 172 different imaging features were extracted, including PET, CT, and spatial probability features. These imaging features were used as inputs into different machine learning algorithms. The impact of different deterministic factors affecting classification performance was assessed. Results . The factors that most impacted classification performance were the machine learning algorithm and the lesion identification method. Random forests (RF) had the highest classification performance. For lesion segmentation, using SORT (AUC = 0.95 [95%CI = 0.94–0.95], sensitivity = 88% [86%–90%], and specificity = 0.89 [0.87–0.90]) resulted in superior classification performance ( p < 0.001) compared to SUV > 10 g ml −1 (AUC = 0.87) and SUV > 15 g ml −1 (AUC = 0.86). While there was only moderate agreement between physicians in lesion classification ( κ = 0.53 [95% CI = 0.52–0.53]), classification performance was high using any of the four physicians as ground truth (AUC range: 0.91–0.93). Conclusion . We have developed the first whole-body automatic disease classification tool for NaF PET using RF, and demonstrated its ability to replicate different physicians’ classification tendencies. This enables fully-automated analysis of whole-body NaF PET/CT images.
- 1 November 2018
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 63 (22) , 225018
A statistically optimized regional thresholding method (SORT) for bone lesion detection in 18 F-NaF PET/CT imaging
Identification of individual lesions on 18 F-NaF PET bone scans is a time-consuming and often subjective process that makes accurate characterization of disease burden challenging. Current automated methods either underestimate disease or struggle with high false positive rates. We developed a statistically optimized regional thresholding (SORT) method that optimizes detection of bone lesions. This study assessed 18 F-NaF PET/CT scans of 37 bone metastatic prostate cancer patients. Each PET image was divided into 19 skeletal regions. Areas of disease in each skeletal region were identified by an experienced nuclear medicine physician. A region of interest (ROI) was placed at each disease location and local maxima were extracted for both healthy and diseased ROIs. Secondary physician review was performed after identification of suspicious local maxima. Region-specific SUV thresholds were determined based on receiver operating characteristic (ROC) analysis optimized for detection of malignant disease. The detection performance of the SORT thresholds were compared to commonly used SUV > 10 g ml −1 (SUV 10 ) and SUV > 15 g ml −1 (SUV 15 ) global thresholds. The sensitivity of the SORT thresholds to various factors was evaluated, such as the number of subjects evaluated or image reconstruction settings. 1751 lesions were manually identified by the nuclear medicine physician. SORT identified different thresholds in each skeletal region (SUV range: 3–13 g ml −1 ). Region-specific SORT thresholding resulted in higher sensitivity (95.8%) than commonly used global thresholds (82.8% for SUV 10 and 58.4% for SUV 15 ) while maintaining a high specificity (97.1%, compared to 97.3% for SUV 10 and 100.0% for SUV 15 ). Factors, such as reconstruction settings, had minimal impact on threshold optimization, resulting in an average change of 10% (range: 2%–17%) in thresholds for each factor. Region-specific SUV thresholding of NaF PET images for bone lesion detection in metastatic prostate patients was found to be superior to current global thresholding methods.
- 1 November 2018
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 63 (22) , 225018
A statistically optimized regional thresholding method (SORT) for bone lesion detection in 18 F-NaF PET/CT imaging
Identification of individual lesions on 18 F-NaF PET bone scans is a time-consuming and often subjective process that makes accurate characterization of disease burden challenging. Current automated methods either underestimate disease or struggle with high false positive rates. We developed a statistically optimized regional thresholding (SORT) method that optimizes detection of bone lesions. This study assessed 18 F-NaF PET/CT scans of 37 bone metastatic prostate cancer patients. Each PET image was divided into 19 skeletal regions. Areas of disease in each skeletal region were identified by an experienced nuclear medicine physician. A region of interest (ROI) was placed at each disease location and local maxima were extracted for both healthy and diseased ROIs. Secondary physician review was performed after identification of suspicious local maxima. Region-specific SUV thresholds were determined based on receiver operating characteristic (ROC) analysis optimized for detection of malignant disease. The detection performance of the SORT thresholds were compared to commonly used SUV > 10 g ml −1 (SUV 10 ) and SUV > 15 g ml −1 (SUV 15 ) global thresholds. The sensitivity of the SORT thresholds to various factors was evaluated, such as the number of subjects evaluated or image reconstruction settings. 1751 lesions were manually identified by the nuclear medicine physician. SORT identified different thresholds in each skeletal region (SUV range: 3–13 g ml −1 ). Region-specific SORT thresholding resulted in higher sensitivity (95.8%) than commonly used global thresholds (82.8% for SUV 10 and 58.4% for SUV 15 ) while maintaining a high specificity (97.1%, compared to 97.3% for SUV 10 and 100.0% for SUV 15 ). Factors, such as reconstruction settings, had minimal impact on threshold optimization, resulting in an average change of 10% (range: 2%–17%) in thresholds for each factor. Region-specific SUV thresholding of NaF PET images for bone lesion detection in metastatic prostate patients was found to be superior to current global thresholding methods.
- Open Access
- 1 October 2016
- Published by Elsevier in European Journal of Radiology
- Vol. 85 (10) , 1744-1749
Limited diagnostic value of Dual-Time-Point 18F-FDG PET/CT imaging for classifying solitary pulmonary nodules in granuloma-endemic regions both at visual and quantitative analyses
- 1 October 2016
- Published by Elsevier in European Journal of Radiology
- Vol. 85 (10) , 1744-1749
Open AccessLimited diagnostic value of Dual-Time-Point 18F-FDG PET/CT imaging for classifying solitary pulmonary nodules in granuloma-endemic regions both at visual and quantitative analyses
- Open Access
- 21 January 2016
- Published by Wiley in BioMed Research International
- Vol. 2016, 1-5
Thyroid Remnant Estimation by Diagnostic Dose Scintigraphy or Scintigraphy after Thyroidectomy: A Comparison with Therapeutic Dose Imaging
In this clinical study, we have compared routine diagnostic dose131I scan andthyroid scintigraphy with therapeutic dose131I imaging for accurate thyroid remnant estimation after total thyroidectomy. We conducted a retrospective review of the patients undergoing total thyroidectomy for differentiated thyroid carcinoma (DTC) and subsequently receiving radioactive iodine (RAI) treatment to ablate remnant thyroid tissue. All patients had therapeutic dose RAI whole body scan, which was compared with that of diagnostic dose RAI,thyroid scan, and ultrasound examination. We concluded that therapeutic dose RAI scan reveals some extent thyroid remnant in all DTC patients following total thyroidectomy. Diagnostic RAI scan is much superior to ultrasound andthyroid scan for the postoperative estimation of thyroid remnant. Ultrasound andthyroid scan provide little information for thyroid remnant estimation and, therefore, would not replace diagnostic RAI scan.- 21 January 2016
- Published by Wiley in BioMed Research International
- Vol. 2016, 1-5
Open AccessThyroid Remnant Estimation by Diagnostic Dose Scintigraphy or Scintigraphy after Thyroidectomy: A Comparison with Therapeutic Dose Imaging
In this clinical study, we have compared routine diagnostic dose131I scan andthyroid scintigraphy with therapeutic dose131I imaging for accurate thyroid remnant estimation after total thyroidectomy. We conducted a retrospective review of the patients undergoing total thyroidectomy for differentiated thyroid carcinoma (DTC) and subsequently receiving radioactive iodine (RAI) treatment to ablate remnant thyroid tissue. All patients had therapeutic dose RAI whole body scan, which was compared with that of diagnostic dose RAI,thyroid scan, and ultrasound examination. We concluded that therapeutic dose RAI scan reveals some extent thyroid remnant in all DTC patients following total thyroidectomy. Diagnostic RAI scan is much superior to ultrasound andthyroid scan for the postoperative estimation of thyroid remnant. Ultrasound andthyroid scan provide little information for thyroid remnant estimation and, therefore, would not replace diagnostic RAI scan.