A Grad-CAM-based knowledge distillation method for the detection of tuberculosis

Automatic screening for tuberculosis (TB) from X-ray images using artificial intelligence techniques has attracted the attention of researchers in the fields of computing and medicine. However, existing models are computationally intensive and require high computer hardware, which limits the use of people in areas where medical resources are scarce. Another problem with the existing model is poor interpretability. The model only provides the final result and lacks intuitive information about the location of the lesion. To solve these problems, this paper proposes a Grad-CAM-based knowledge distillation method for the detection of TB. Firstly, this study used Unet to extract the lung region, avoiding the influence of regions outside the lung on the detection results. Subsequently, five models (Densenet121, Inception V3, Resnet18, Mobilenet V3, VGG16) are applied to TB detection, and the attention maps of each model are visualized using Grad-CAM. These attention maps are applied to knowledge distillation to finally obtain a lightweight interpretable TB detection model. This model achieves 91.2% and 85.7% accuracy on Shenzhen and Montgomery datasets, which verifies the effectiveness of the model.