Background-Aim: Artificial intelligence techniques (machine learning, deep learning, and radiomics) play an increasing role in nuclear medicine to diagnose oncological diseases. Radiomics provides the possibility to build classification and/or regression models based on quantitative features extracted from positron emission tomography (PET) scan data. The aim of our work was the discrimination between benign and malignant solitary pulmonary nodules (SPN), as it may provide an in vivo classification of the disease, thus avoiding invasive diagnostic techniques. Methods: Baseline PET/CT scans (Discovery 710, GE Healthcare) of 83 histologically-confirmed lung nodules (44 adenocarcinomas, 39 benign nodules) from as many patients (46 males, 37 females, age = 67.2 ± 9.0 [44–83] year) undergoing examination at the Unit of Nuclear Medicine of the University of Sassari, (Italy), between November 2014 and July 2019 were retrospectively analysed. For the radiomics analysis we considered 5 conventional features, 8 firstorder texture features, 9 s-order texture features and 6 shape features computed on the PET and CT signal separately for a total of 42 features. A subset of 25 features that showed statistically-significant differences (determined via Mann–Whitney U test) between the two phenotypes were eventually retained for the classification step. Four classification models (k-NN, Logistic Regression, Random Forest and Support Vector Machines) were fed with the selected features to test their ability to discriminate the adenocarcinomas from the benign lesions. Accuracy estimation was based on split-sample validation with stratified sampling at 70% train rate, and the results were averaged over 250 random splits into train and test set for a stable estimation. Results: The performance of the four classification models was very close; the best classifier achieved an accuracy of 80.2% (sensitivity = 75.1%, specificity = 94.4%). Conclusions: Our results suggest that adenocarcinoma has a radiomics signature that allows to differentiate it from benign nodules, therefore representing an ‘‘identity card’’ of this type of lung cancer. This confirms the potential benefits of radiomics in the management of patients with indeterminate pulmonary nodules. This work is however not exempt from limitations, in particular the relatively contained sample size and the retrospective nature of the study. The results should be further validated in larger, ideally prospective studies.
Radiomics analysis on PET/CT scans to discriminate malignant (adenocarcinoma) and benign pulmonary nodules / Rondini, M.; Bianconi, F.; Fravolini, M. L.; Minestrini, M.; Stazza, M. L.; Filippi, L.; Marongiu, A.; Nuvoli, S.; Spanu, A.; Palumbo, B.. - In: CLINICAL AND TRANSLATIONAL IMAGING. - ISSN 2281-5872. - (2022).
Radiomics analysis on PET/CT scans to discriminate malignant (adenocarcinoma) and benign pulmonary nodules
M. Rondini;M. L. Stazza;A. Marongiu;S. Nuvoli;A. Spanu;
2022-01-01
Abstract
Background-Aim: Artificial intelligence techniques (machine learning, deep learning, and radiomics) play an increasing role in nuclear medicine to diagnose oncological diseases. Radiomics provides the possibility to build classification and/or regression models based on quantitative features extracted from positron emission tomography (PET) scan data. The aim of our work was the discrimination between benign and malignant solitary pulmonary nodules (SPN), as it may provide an in vivo classification of the disease, thus avoiding invasive diagnostic techniques. Methods: Baseline PET/CT scans (Discovery 710, GE Healthcare) of 83 histologically-confirmed lung nodules (44 adenocarcinomas, 39 benign nodules) from as many patients (46 males, 37 females, age = 67.2 ± 9.0 [44–83] year) undergoing examination at the Unit of Nuclear Medicine of the University of Sassari, (Italy), between November 2014 and July 2019 were retrospectively analysed. For the radiomics analysis we considered 5 conventional features, 8 firstorder texture features, 9 s-order texture features and 6 shape features computed on the PET and CT signal separately for a total of 42 features. A subset of 25 features that showed statistically-significant differences (determined via Mann–Whitney U test) between the two phenotypes were eventually retained for the classification step. Four classification models (k-NN, Logistic Regression, Random Forest and Support Vector Machines) were fed with the selected features to test their ability to discriminate the adenocarcinomas from the benign lesions. Accuracy estimation was based on split-sample validation with stratified sampling at 70% train rate, and the results were averaged over 250 random splits into train and test set for a stable estimation. Results: The performance of the four classification models was very close; the best classifier achieved an accuracy of 80.2% (sensitivity = 75.1%, specificity = 94.4%). Conclusions: Our results suggest that adenocarcinoma has a radiomics signature that allows to differentiate it from benign nodules, therefore representing an ‘‘identity card’’ of this type of lung cancer. This confirms the potential benefits of radiomics in the management of patients with indeterminate pulmonary nodules. This work is however not exempt from limitations, in particular the relatively contained sample size and the retrospective nature of the study. The results should be further validated in larger, ideally prospective studies.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.