Model interpretation

We demonstrated pulmonary nodules in the form of a heatmap display on each chest radiograph for easy visualization, and we presented the positive probability score as an index value (0.0-1.0), which indicated the possibility of pulmonary nodules using class activation maps (CAMs)¹⁶⁾. To generate the CAMs, we fed an image into the fully trained network and extracted the feature maps that were output by the final convolutional layer. With f_k as the k^th feature map and w_c,k as the weight in the final classification layer for feature map k leading to pulmonary nodules, we obtained a map M_c of the most salient features, which were used to classify the images as having pulmonary nodules by taking the weighted sum of the feature maps using their associated weights. The equation is as follows:

M _c = Σ_k w_c,k · f_k

We identified the most important features used by the model to predict the presence of pulmonary nodules by upscaling the map M_c to the dimensions of the image and overlaying the image. Our novel AI system underwent mechanical learning of training data, which were obtained using the same radiographic apparatus to eliminate the effects of differences between equipment.