A new study has developed and tested a deep learning-based system for the automatic detection and classification of lung nodules using computed tomography (CT) images. The research, conducted on 82 patients from the publicly available LIDC-IDRI database, addresses challenges in early lung cancer diagnosis by applying a convolutional neural network (CNN) architecture.
Lung cancer is the leading cause of cancer-related deaths worldwide. Early identification of pulmonary nodules is essential for timely intervention. However, traditional computer-aided detection methods have been limited by high false-positive rates and low sensitivity.
The study’s method involves several steps: image preprocessing, segmentation of lung parenchyma with Otsu's thresholding and morphological operations, candidate nodule detection, feature extraction, and final classification through a CNN model. The network architecture includes two convolutional layers with 20 and 30 filters respectively, ReLU activation functions, max-pooling layers, and a Softmax output layer. Training was performed using mini-batches of 32 images over 50 epochs with Stochastic Gradient Descent with Momentum optimization.
Results showed that the CNN-based system detected pulmonary nodules effectively and classified them as benign or malignant with high accuracy. On the LIDC-IDRI dataset, it achieved a sensitivity of 98.7%, specificity of 97.5%, precision of 97.9%, and overall accuracy of 98.4%. The model’s performance was compared to recent approaches such as hybrid CNN-long short-term memory models and ResNet-based models; it delivered similar results while requiring less computational power.
The researchers also evaluated the system’s ability to distinguish between different types of nodules—solid, partially frosted, and totally frosted—with satisfactory results.
"The proposed CNN-based system demonstrates the feasibility and robustness of deep learning for automatic lung nodule detection and classification," the study states. "Despite strong results, the study acknowledges limitations such as single-database validation and a relatively small training size. Future work will focus on validating the model across other datasets (e.g., ELCAP, NELSON) and optimizing multi-class classification performance to enhance generalizability and clinical applicability."
The authors note that further testing on additional datasets will be necessary to confirm these findings in broader clinical settings.
