Pulmonary Medical Image Recognition Based on Deep Learning

Pulmonary nodules are important indicators for the early diagnosis of lung can- cer, and their identification and classification are of great significance. At present, there exist significant domain discrepancies between source and target datasets when using transfer learning-based recognition algorithms, which leads to poor feature extraction of pulmonary nodules and thus unsatisfactory results. Therefore, this paper proposes an improved convolutional neural network model based on a mod- ified neural network architecture. The model integrates features extracted by the pre-trained GoogLeNet Inception V3 network to enhance its ability to extract relev- ant features. In order to determine the optimal combination, different groups were tested using accuracy as the evaluation metric. Experiments were conducted on the LUNA16 lung nodule dataset. The results from cross-validation testing show that the improved network achieves an accuracy of 88.80% and a sensitivity of 87.15%. In terms of recognition accuracy and sensitivity, the proposed method outperforms GoogLeNet Inception V3, with improvements of 2.72 and 2.19 percentage points, respectively. Even when tested on small-scale datasets, the model demonstrates bet- ter generalization capability. This method can provide objective reference indicators for clinical diagnosis.