Margarita Kirienko1, Luca Cozzi1,2, Alexia Rossi1,3, Emanuele Voulaz4, Lidija Antunovic5, Antonella Fogliata2, Arturo Chiti1,5, Martina Sollini6. 1. Department of Biomedical Sciences, Humanitas University, Via Rita Levi Montalcini 4, Pieve Emanuele, 20090, Milan, Italy. 2. Radiotherapy and Radiosurgery, Humanitas Clinical and Research Hospital, Rozzano, Milan, Italy. 3. Radiology, Humanitas Clinical and Research Hospital, Rozzano, Milan, Italy. 4. Thoracic Surgery, Humanitas Clinical and Research Hospital, Rozzano, Milan, Italy. 5. Nuclear Medicine, Humanitas Clinical and Research Hospital, Rozzano, Milan, Italy. 6. Department of Biomedical Sciences, Humanitas University, Via Rita Levi Montalcini 4, Pieve Emanuele, 20090, Milan, Italy. martina.sollini@cancercenter.humanitas.it.
Abstract
PURPOSE: To evaluate the ability of CT and PET radiomics features to classify lung lesions as primary or metastatic, and secondly to differentiate histological subtypes of primary lung cancers. METHODS: A cohort of 534 patients with lung lesions were retrospectively studied. Radiomics texture features were extracted using the LIFEx package from semiautomatically segmented PET and CT images. Histology data were recorded in all patients. The patient cohort was divided into a training and a validation group and linear discriminant analysis (LDA) was performed to classify the lesions using both direct and backward stepwise methods. The robustness of the procedure was tested by repeating the entire process 100 times with different assignments to the training and validation groups. Scoring metrics included analysis of the receiver operating characteristic curves in terms of area under the curve (AUC), sensitivity, specificity and accuracy. RESULTS: Radiomics features extracted from CT and PET datasets were able to differentiate primary tumours from metastases in both the training and the validation group (AUCs 0.79 ± 0.03 and 0.70 ± 0.04, respectively, from the CT dataset; AUCs 0.92 ± 0.01 and 0.91 ± 0.03, respectively, from the PET dataset). The AUC cut-off thresholds identified by LDA using direct and backward elimination strategies were -0.79 ± 0.06 and -0.81 ± 0.08, respectively (CT dataset) and -0.69 ± 0.05 and -0.68 ± 0.04, respectively (PET dataset). For differentiation between primary subgroups based on CT features, the AUCs in the training and validation groups were 0.81 ± 0.02 and 0.69 ± 0.04 for adenocarcinoma (Adc) vs. squamous cell carcinoma (Sqc) or "Other", 0.85 ± 0.02 and 0.70 ± 0.05 for Sqc vs. Adc or Other, and 0.77 ± 0.03 and 0.57 ± 0.05 for Other vs. Adc or Sqc. The same analyses for the PET data revealed AUCs of 0.90 ± 0.10 and 0.80 ± 0.04, 0.80 ± 0.02 and 0.61 ± 0.06, and 0.97 ± 0.01 and 0.88 ± 0.04, respectively. CONCLUSION: PET radiomics features were able to differentiate between primary and metastatic lung lesions and showed the potential to identify primary lung cancer subtypes.
PURPOSE: To evaluate the ability of CT and PET radiomics features to classify lung lesions as primary or metastatic, and secondly to differentiate histological subtypes of primary lung cancers. METHODS: A cohort of 534 patients with lung lesions were retrospectively studied. Radiomics texture features were extracted using the LIFEx package from semiautomatically segmented PET and CT images. Histology data were recorded in all patients. The patient cohort was divided into a training and a validation group and linear discriminant analysis (LDA) was performed to classify the lesions using both direct and backward stepwise methods. The robustness of the procedure was tested by repeating the entire process 100 times with different assignments to the training and validation groups. Scoring metrics included analysis of the receiver operating characteristic curves in terms of area under the curve (AUC), sensitivity, specificity and accuracy. RESULTS: Radiomics features extracted from CT and PET datasets were able to differentiate primary tumours from metastases in both the training and the validation group (AUCs 0.79 ± 0.03 and 0.70 ± 0.04, respectively, from the CT dataset; AUCs 0.92 ± 0.01 and 0.91 ± 0.03, respectively, from the PET dataset). The AUC cut-off thresholds identified by LDA using direct and backward elimination strategies were -0.79 ± 0.06 and -0.81 ± 0.08, respectively (CT dataset) and -0.69 ± 0.05 and -0.68 ± 0.04, respectively (PET dataset). For differentiation between primary subgroups based on CT features, the AUCs in the training and validation groups were 0.81 ± 0.02 and 0.69 ± 0.04 for adenocarcinoma (Adc) vs. squamous cell carcinoma (Sqc) or "Other", 0.85 ± 0.02 and 0.70 ± 0.05 for Sqc vs. Adc or Other, and 0.77 ± 0.03 and 0.57 ± 0.05 for Other vs. Adc or Sqc. The same analyses for the PET data revealed AUCs of 0.90 ± 0.10 and 0.80 ± 0.04, 0.80 ± 0.02 and 0.61 ± 0.06, and 0.97 ± 0.01 and 0.88 ± 0.04, respectively. CONCLUSION: PET radiomics features were able to differentiate between primary and metastatic lung lesions and showed the potential to identify primary lung cancer subtypes.
Authors: S J Swensen; M D Silverstein; E S Edell; V F Trastek; G L Aughenbaugh; D M Ilstrup; C D Schleck Journal: Mayo Clin Proc Date: 1999-04 Impact factor: 7.616
Authors: Hendra Budiawan; Gi Jeong Cheon; Hyung-Jun Im; Soo Jin Lee; Jin Chul Paeng; Keon Wook Kang; June-Key Chung; Dong Soo Lee Journal: Nucl Med Mol Imaging Date: 2013-08-21
Authors: Paul Flechsig; Philipp Frank; Clemens Kratochwil; Gerald Antoch; Daniel Rath; Jan Moltz; Michael Rieser; Arne Warth; Hans-Ulrich Kauczor; Lawrence H Schwartz; Uwe Haberkorn; Frederik L Giesel Journal: Mol Imaging Biol Date: 2017-04 Impact factor: 3.488
Authors: P P Leong; B Rezai; W M Koch; A Reed; D Eisele; D J Lee; D Sidransky; J Jen; W H Westra Journal: J Natl Cancer Inst Date: 1998-07-01 Impact factor: 13.506