Objective: To enhance the positive predictive value (PPV) of chest digital tomosynthesis (DTS) in the lung cancer detection with the analysis of radiomics features. Method: The investigation was carried out within the SOS clinical trial (NCT03645018) for lung cancer screening with DTS. Lung nodules were identified by visual analysis and then classified using the diameter and the radiological aspect of the nodule following lung-RADS. Haralick texture features were extracted from the segmented nodules. Both semantic variables and radiomics features were used to build a predictive model using logistic regression on a subset of variables selected with backward feature selection and using two machine learning: a Random Forest and a neural network with the whole subset of variables. The methods were applied to a train set and validated on a test set where diagnostic accuracy metrics were calculated. Results: Binary visual analysis had a good sensitivity (0.95) but a low PPV (0.14). Lung-RADS classification increased the PPV (0.19) but with an unacceptable low sensitivity (0.65). Logistic regression showed a mildly increased PPV (0.29) but a lower sensitivity (0.20). Random Forest demonstrated a moderate PPV (0.40) but with a low sensitivity (0.30). Neural network demonstrated to be the best predictor with a high PPV (0.95) and a high sensitivity (0.90). Conclusions: The neural network demonstrated the best PPV. The use of visual analysis along with neural network could help radiologists to reduce the number of false positive in DTS. Key Points: • We investigated several approaches to enhance the positive predictive value of chest digital tomosynthesis in the lung cancer detection. • Neural network demonstrated to be the best predictor with a nearly perfect PPV. • Neural network could help radiologists to reduce the number of false positive in DTS.

Artificial intelligence and radiomics enhance the positive predictive value of digital chest tomosynthesis for lung cancer detection within SOS clinical trial

Berchialla P.;
2020

Abstract

Objective: To enhance the positive predictive value (PPV) of chest digital tomosynthesis (DTS) in the lung cancer detection with the analysis of radiomics features. Method: The investigation was carried out within the SOS clinical trial (NCT03645018) for lung cancer screening with DTS. Lung nodules were identified by visual analysis and then classified using the diameter and the radiological aspect of the nodule following lung-RADS. Haralick texture features were extracted from the segmented nodules. Both semantic variables and radiomics features were used to build a predictive model using logistic regression on a subset of variables selected with backward feature selection and using two machine learning: a Random Forest and a neural network with the whole subset of variables. The methods were applied to a train set and validated on a test set where diagnostic accuracy metrics were calculated. Results: Binary visual analysis had a good sensitivity (0.95) but a low PPV (0.14). Lung-RADS classification increased the PPV (0.19) but with an unacceptable low sensitivity (0.65). Logistic regression showed a mildly increased PPV (0.29) but a lower sensitivity (0.20). Random Forest demonstrated a moderate PPV (0.40) but with a low sensitivity (0.30). Neural network demonstrated to be the best predictor with a high PPV (0.95) and a high sensitivity (0.90). Conclusions: The neural network demonstrated the best PPV. The use of visual analysis along with neural network could help radiologists to reduce the number of false positive in DTS. Key Points: • We investigated several approaches to enhance the positive predictive value of chest digital tomosynthesis in the lung cancer detection. • Neural network demonstrated to be the best predictor with a nearly perfect PPV. • Neural network could help radiologists to reduce the number of false positive in DTS.
1
7
Artificial intelligence; Lung cancer; Thoracic radiography
Chauvie S.; De Maggi A.; Baralis I.; Dalmasso F.; Berchialla P.; Priotto R.; Violino P.; Mazza F.; Melloni G.; Grosso M.; Biggi A.; Campione A.; Fortunato M.; Colantonio I.; Stanzi A.; Noceti P.; Pellegrino P.; Russi E.
File in questo prodotto:
File Dimensione Formato  
Radiomics-SOS_rev2_pb.pdf

Accesso aperto

Tipo di file: PREPRINT (PRIMA BOZZA)
Dimensione 211.55 kB
Formato Adobe PDF
211.55 kB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/2318/1739252
Citazioni
  • ???jsp.display-item.citation.pmc??? 3
  • Scopus 10
  • ???jsp.display-item.citation.isi??? 9
social impact