SPECT bone scan image classification by fusing multi-attention mechanism with deep residual networks

SPECT nuclear medicine imaging is an effective means of clinically diagnosing bone metastatic diseases. This paper proposes a ResNet improvement model combining Convolutional Block Attention Module (CBAM) and contextual transformer (CoT) attention mechanism to achieve accurate classification of SPECT images. The improved model incorporates the CoTattention mechanism into the residual module of the network to effectively utilize the global information of image features. A CBAM module, which combines channel and spatial attention, is added before the fully connected layer to extract useful information from image features more effectively. Thus, a classification model capable of accurately identifying bone metastatic lesion features is constructed. Experimental results on a real SPECT image dataset show that the proposed model performs well in the classification of SPECT images, obtaining values of 95.84%, 96.12%, 95.88%, and 95.83 % in Accuracy, Precision, Recall, and F-1 Score evaluation metrics, respectively. Considering the assessment of model effectiveness by each evaluation index, we prefer the result of F-1 Score as the main one, and further comparison experiments with several similar models prove the effectiveness of the model proposed in this paper.