Research on Cancer Recognition Based on Deep Transfer Learning and CT Images

Artificial intelligence technology has been applied in real life to help people solve various problems, such as smart transportation, smart cities, and smart healthcare. Machine learning, as one of the artificial intelligence technologies, has been widely used in various medical image diagnosis and recognition, including CT or X-ray-based pneumonia identification, cancer identification, and more. However, traditional machine learning algorithms such as random forest, support vector machine, and shallow neural networks cannot fully extract image information, resulting in low recognition accuracy and requiring manual feature extraction, which increases the model’s cost. Additionally, most existing deep learning algorithms, such as fully connected DNN and convolutional neural networks CNN, require a large number of images to train the model. However, obtaining this image information in real life is costly. Therefore, this article proposes a deep transfer learning algorithm for CT cancer image recognition, based on MobileNet. Specifically, we freeze the shallow parameters of MobileNet that have already been trained; then, we fine-tune the deep network parameters to achieve cancer recognition based on CT images. Unlike traditional deep learning algorithms, we use a GlobalAveragePooling layer before the output layer to obtain more critical hidden information while reducing the model’s parameters. Compared with traditional deep learning algorithms, the proposed deep transfer learning algorithm can obtain satisfactory diagnostic results with less training data. Compared to starting training from MobileNet, it saves a lot of training time while maintaining high accuracy. We designed, trained, and experimentally verified the effectiveness of the proposed algorithm on a real dataset, achieving an accuracy rate of 88.42%.