Integrating Transfer Learning and Feature Aggregation into Self-defined Convolutional Neural Network for Automated Detection of Lung Cancer Bone Metastasis

Purpose: To improve the diagnosis accuracy and efficiency of lung cancer bone metastasis routinely performed by nuclear medicine physicians, we propose a deep learning-based image classification model that can learn the features from two views of an image first, then aggregate them, and finally classify the image into the presence or absence of bone metastasis. Methods: We present a new network that can automatically classify scintigraphy images collected from the clinical diagnosis of metastasis in patients with lung cancer. The proposed network consists of pre-training, transfer learning, and two-view feature aggregation. In the pre-training stage, the proposed model is trained on a source dataset of Chest X-Ray. In the transfer learning stage, the pre-trained model is fine-turned on the target dataset of scintigraphy images. The extracted features from anterior and posterior views of an image are aggregated in the final stage. The classification network can detect the presence or absence of metastases in scintigraphy images. Results: Experimental evaluations on a set of clinical scintigraphy images showed that the proposed network performed well for automatically classifying metastatic images with the mean scores of 0.7710, 0.8311, 0.6827 and 0.7475 on the accuracy, precision, recall and F-1 score, respectively. Conclusion: The proposed classification network can predict whether an image shows lung cancer-caused metastasis with state-of-the-art performance.