M Colevray1, V M Tatard-Leitman2, S Gouttard1, P Douek3, L Boussel4. 1. Department of radiology, hôpital de la Croix-Rousse, 103, Grande rue de la Croix-Rousse, 69004 Lyon, France. 2. Unité CNRS UMR 5220, CREATIS, Inserm U1206, Insa Lyon, université Lyon 1, université Jean-Monnet Saint-Étienne, 7, avenue Jean-Capelle, 69100 Villeurbanne, France; Department of radiology, Louis-Pradel hospital, 59, boulevard Pinel, 69500 Bron, France. 3. Department of radiology, hôpital de la Croix-Rousse, 103, Grande rue de la Croix-Rousse, 69004 Lyon, France; Department of radiology, Louis-Pradel hospital, 59, boulevard Pinel, 69500 Bron, France. 4. Department of radiology, hôpital de la Croix-Rousse, 103, Grande rue de la Croix-Rousse, 69004 Lyon, France; Unité CNRS UMR 5220, CREATIS, Inserm U1206, Insa Lyon, université Lyon 1, université Jean-Monnet Saint-Étienne, 7, avenue Jean-Capelle, 69100 Villeurbanne, France. Electronic address: loic.boussel@chu-lyon.fr.
Abstract
INTRODUCTION: The purpose of this study was to develop a convolutional neural network (CNN) to determine the extent of over-scanning in the Z-direction associated with lung computed tomography (CT) examinations. MATERIALS AND METHODS: The CT examinations of 250 patients were used to train the machine learning software and 100 were used to validate the results. Each lung CT examination was divided into cervical, lung, and abdominal areas by the CNN and 2 independent radiologists, and the length of each area was measured. Every part above or below the lung marks was labeled as over-scanning. The accuracy of the CNN was calculated after the training phase and agreement between CNN and radiologists was assessed using kappa statistics during the validation phase. After validation the software was used to estimate the length of each of the three areas and the total over-scanning in further 1000 patients. RESULTS: An accuracy of 0.99 was found for the testing dataset and a very good agreement (kappa=0.98) between the CNN and the radiologists' evaluation was found for the validation dataset. Over-scanning was 22.8% with the CNN and 22.2% with the radiologists. The degree of over-scanning was 22.6% in 1000 lung CT examinations. CONCLUSION: Our study shows a substantial over estimation of the length of the area to be scanned during lung CT and thus an unnecessary patient's over-exposure to ionizing radiation. This over-scanning can be assessed easily, reliably and quickly using CNN.
INTRODUCTION: The purpose of this study was to develop a convolutional neural network (CNN) to determine the extent of over-scanning in the Z-direction associated with lung computed tomography (CT) examinations. MATERIALS AND METHODS: The CT examinations of 250 patients were used to train the machine learning software and 100 were used to validate the results. Each lung CT examination was divided into cervical, lung, and abdominal areas by the CNN and 2 independent radiologists, and the length of each area was measured. Every part above or below the lung marks was labeled as over-scanning. The accuracy of the CNN was calculated after the training phase and agreement between CNN and radiologists was assessed using kappa statistics during the validation phase. After validation the software was used to estimate the length of each of the three areas and the total over-scanning in further 1000 patients. RESULTS: An accuracy of 0.99 was found for the testing dataset and a very good agreement (kappa=0.98) between the CNN and the radiologists' evaluation was found for the validation dataset. Over-scanning was 22.8% with the CNN and 22.2% with the radiologists. The degree of over-scanning was 22.6% in 1000 lung CT examinations. CONCLUSION: Our study shows a substantial over estimation of the length of the area to be scanned during lung CT and thus an unnecessary patient's over-exposure to ionizing radiation. This over-scanning can be assessed easily, reliably and quickly using CNN.