Lung cancer continues to rank as the top cause of cancer deaths globally. Detecting pulmonary nodules early plays a vital role in diagnosis and treatment success. Traditional computer-aided systems often struggle with high false positives and poor sensitivity, but deep learning innovations like convolutional neural networks (CNNs) offer promising improvements in accuracy.
Study Overview and Dataset
Researchers developed and tested an automated CNN-based approach for identifying and classifying lung nodules using computed tomography (CT) scans from the LIDC-IDRI database. The analysis covered 82 patients, encompassing 10,496 CT slices.
Detection and Classification Process
The system follows five key steps: image preprocessing, lung tissue segmentation via Otsu’s thresholding and morphological operations, nodule candidate identification, feature extraction, and CNN-based classification. The CNN features two convolutional layers with 20 and 30 filters using 3×3 kernels, ReLU activation, max-pooling layers, and a Softmax output. Training used mini-batches of 32 over 50 epochs with Stochastic Gradient Descent and Momentum (learning rate 0.001, momentum 0.9).
Performance Metrics
The model delivers strong results in nodule detection and distinguishing benign from malignant cases. On the LIDC-IDRI dataset, it records 98.7% sensitivity, 97.5% specificity, 97.9% precision, and 98.4% overall accuracy. It outperforms or matches hybrid CNN-LSTM and ResNet models while requiring less computational power. Subtype classification for solid, partially frosted, and totally frosted nodules also proves effective.
Limitations and Next Steps
While results show robustness, challenges include reliance on a single database and limited training data. Ongoing efforts aim to test the model on datasets like ELCAP and NELSON, alongside refining multi-class classification for broader clinical use.
