Improved Convolutional Neural Network Lung Cancer Classification Detection Method Based on Transfer Learning and Model Compression

Lung cancer, one of the most life-threatening malignancies, has witnessed a significant increase in its incidence and mortality rates over the past few decades. Accurate determination of the histological type of lung cancer is crucial for developing appropriate treatment strategies. However, traditional medical imaging analysis, such as magnetic resonance imaging and computed tomography (CT), faces challenges due to sensor noise, patient variations, and pathological states, which may lead to misinterpretation and errors. In this regard, deep learning approaches can assist in rapidly distinguishing different types of lung cancer by learning complex patterns and features from medical imaging data, aiding clinicians in precise diagnosis. This research aims to address the limitations of existing lung cancer classification models, including suboptimal accuracy, slow inference speed, high computational and storage requirements, and vulnerability to overfitting. To achieve this, we propose an improved lung cancer classification and detection method based on deep learning. Our approach leverages transfer learning by utilizing pre-trained models, data augmentation techniques, and model compression methods, including global unstructured pruning and model distillation. Experimental results demonstrate that the teacher model achieves 100% accuracy, recall, precision, and F1-score on the dataset, with a loss of 0.00007575038. The distilled student model, after pruning 55.15% of its parameters, achieves a 52.5% reduction in size. It also attains 100% accuracy, recall, precision, and F1-score, with an average loss of 0.004158932.