|
Deep learning-based classification of lung nodules on CT images: A multicenter study of early lung cancer detection
|
C. Qi   , Q. Jiang |
| Baicheng Central Hospital, Baicheng, Jilin, 137000, China , 77073199@bcmc.edu.cn |
|
|
Abstract: (7 Views) |
Background: Early lung cancer detection relies on accurate characterization of pulmonary nodules on computed tomography (CT). However, variability in CT acquisition and radiological interpretation may reduce reproducibility across institutions. This study evaluated a deep learning (DL) framework for benign-malignant lung nodule classification using multicenter CT datasets. Materials and Methods: Publicly available anonymized CT datasets from LUNA16 and LIDC-IDRI were used. Images were rescaled to a uniform resolution, denoised with Gaussian filtering, and normalized by Min-Max scaling. Nodule regions were segmented using a U-Net encoder-decoder network. Gray-Level Co-Occurrence Matrix (GLCM) texture features were extracted and combined with latent representations learned by an Effective Beetle Antennae Search-driven Variational Autoencoder (EBAS-VAE). Model performance was assessed using accuracy, precision, recall, F1-score, true-positive rate (TPR), false-positive rate (FPR), false discovery rate (FDR), and area under the receiver operating characteristic curve (AUC). Results: On LUNA16, EBAS-VAE achieved accuracy 0.9838, precision 0.94, recall 0.89, F1-score 0.96, and AUC 0.98. On LIDC-IDRI, accuracy was 0.97, TPR 0.97, FPR 0.04, and FDR 0.0001. EBAS-VAE outperformed 3D-CNN and SEOA-DBN comparators. Conclusion: The proposed EBAS-VAE pipeline improved lung nodule classification and reduced false-positive detection in multicenter CT images, supporting its potential as a radiologist-assistive screening tool. |
|
| |
|
|
Full-Text [PDF 908 kb]
(3 Downloads)
|
|
Type of Study: Original Research |
Subject:
Radiation Biology
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
