Convolutional Neural Networks for Multi-scale Lung Nodule Classification in CT: Influence of Hyperparameter Tuning on Performance
Convolutional Neural Networks for Multi-scale Lung Nodule Classification in CT: Influence of Hyperparameter Tuning on Performance
Author(s): Jorge Hernández-Rodríguez, Francisco-Javier Cabrero-Fraile, María-José Rodríguez-CondeSubject(s): Health and medicine and law, ICT Information and Communications Technologies
Published by: UIKTEN - Association for Information Communication Technology Education and Science
Keywords: Convolutional Neural Network; lung nodule; CAD; classification; LIDC-IDRI;
Summary/Abstract: In this study, a system based in Convolutional Neural Networks for differentiating lung nodules and non-nodules in Computed Tomography is developed. Multi-scale patches, extracted from LIDC-IDRI database, are used to train different CNN models. Adjustable hyperparameters are modified sequentially, to study their influence, evaluate learning process and find each size best performing network. Classification accuracies obtained are superior to 87% for all sizes with areas under Receiver Operating Characteristic in the interval (0.936-0.951). Trained models are tested with nodules from an independent database, providing sensitivities above 96%. Performance of trained models is similar to other published articles and show good classification capacities. As a basis for developing CAD systems, recommendations regarding hyperparameter tuning are provided.
Journal: TEM Journal
- Issue Year: 11/2022
- Issue No: 1
- Page Range: 297-306
- Page Count: 10
- Language: English