Developing accurate methods for differentiating benign vs. malignant pulmonary nodules on CT is crucial for the correct management of patients referred for suspicious lung cancer. In this context deep learning by convolutional neural networks (CNN) has been gaining ground as an alternative to conventional methods based on feature engineering, although the use of CNN is often hampered by the lack of sufficiently large datasets for training. Herein we explore the effectiveness of deep features from pre-trained convolutional networks ‘off-the-shelf’ to discriminate benign vs. malignant lung nodules from CT images. To this end we compare three approaches (two classic, one novel) for pseudo-colour image generation which allow the grey-scale CT data to be fed into CNN models designed for and trained on colour images. The classification performance was estimated by internal and cross-validation using two independent datasets (LIDC-IDRI and LUNGx), giving four experimental conditions altogether. Conventional radiomics features were used as baseline reference. The best accuracy achieved by deep features in the four experimental conditions was respectively 88.5%, 69.6%, 68.9% and 65.2%, that obtained through conventional features 85.5%, 60.9%, 65.2% and 63.8%.
Classification of lung nodules on CT via pseudo-colour images and deep features from pre-trained convolutional networks
Francesco Bianconi
;Mario Luca Fravolini;Muhammad Usama Khan;Barbara Palumbo
2024
Abstract
Developing accurate methods for differentiating benign vs. malignant pulmonary nodules on CT is crucial for the correct management of patients referred for suspicious lung cancer. In this context deep learning by convolutional neural networks (CNN) has been gaining ground as an alternative to conventional methods based on feature engineering, although the use of CNN is often hampered by the lack of sufficiently large datasets for training. Herein we explore the effectiveness of deep features from pre-trained convolutional networks ‘off-the-shelf’ to discriminate benign vs. malignant lung nodules from CT images. To this end we compare three approaches (two classic, one novel) for pseudo-colour image generation which allow the grey-scale CT data to be fed into CNN models designed for and trained on colour images. The classification performance was estimated by internal and cross-validation using two independent datasets (LIDC-IDRI and LUNGx), giving four experimental conditions altogether. Conventional radiomics features were used as baseline reference. The best accuracy achieved by deep features in the four experimental conditions was respectively 88.5%, 69.6%, 68.9% and 65.2%, that obtained through conventional features 85.5%, 60.9%, 65.2% and 63.8%.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.