Adaptive morphology aided 2-pathway convolutional neural network for lung nodule classification. (February 2022)

• Record Type: Journal Article
• Authors: Halder, Amitava; Chatterjee, Saptarshi; Dey, Debangshu
• Is Part Of: Biomedical signal processing and control. Volume 72(2022)Part B
• Publication Date: 2022-02
• Subjects: Adaptive morphology -- Deep learning -- Gabor filter -- Computer-aided diagnosis -- Nodule classification -- Lung cancer; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103347
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Early-stage detection and identification of malignant pulmonary nodules can allow proper medication and increase the survival rate of lung cancer patients. High-Resolution Computed Tomography (HRCT) image slices are in use for the screening of lung cancer. However, appropriate identification of lung nodules at the early stage of the disease is challenging owing to similar morphological properties of benign and malignant nodules. Introduction of computer vision and advanced image analysis techniques for the development of Computer-aided diagnosis (CADx) systems have significantly improved the classification performance and increase the speed the interpreting lung CT images for the identification of lung cancer. Deep learning-based techniques have recently emerged as an efficient tool for the improved characterization of lung nodules. In this research work, a deep learning (DL) based framework has been introduced using the concept of adaptive morphology-based operations combined with Gabor filter (GF) for accurate lung nodule classification. The new framework, 2-Pathway Morphology-based Convolutional Neural Network (2PMorphCNN) with its two trainable paths can capture both textural and morphological features of the lung nodules that results in better classification accuracy. The proposed system has been trained and evaluated on LIDC-IDRI dataset and achieved sensitivity, specificity, accuracy of 96.85%, 95.17%, and 96.10% with an Area under the ROC Curve (AUC) of 0.9936 for lung nodule characterization. It has been observed that the reported automatic lung nodule classification framework outperforms other state-of-the-art nodule classification methodologies by capturing and combining textural and morphological features from the HRCT lung nodule image.