Multi-disease classification of whole-body scintigraphy images based on deep learning

Bone scan images has important clinical application value for the early diagnosis and treatment of skeletal lesions in the whole-body. In order to realize classification of multiple diseases (i.e., normal, bone metastases, arthritis, and thyroid cancer) in whole-body bone scintigraphy images, In this paper, the multi-disease classification method based on convolutional neural network was studied on bone scan images, and a self-defined classification model was proposed, which was named SPNT9. The model consists of nine convolutional layers, five pooling layers, two fully connected layers and one softmax layer. Finally, the model was validated with real SPECT images collected. By comparing with VGGNet, ResNet and DenseNet models, The SPNT9 model obtained the best classification performance in the test samples, with Accuracy, Precision, Recall and F-1 of 0.8814, 0.7806, 0.7900 and 0.7899, respectively. The results show that the proposed model can be used to accurately classify multiple diseases in SPECT images.