Fig. 6: Visualization of attention mechanism and interpretability. | Nature Communications

Fig. 6: Visualization of attention mechanism and interpretability.

From: AI-assisted cervical cytology precancerous screening for high-risk population in resource-limited regions using a compact microscope

Fig. 6

Cell-level ROI and instance-level ROI are acquired by the attention score of first-stage model and second-stage model. We frame out instances whose positive probability is in the top 200 as instance-level ROI in a view of a slide using a yellow square (i.e., a and b) and plot the heat map of our attention score in stage one. In addition, class activation maps (CAM) generated from the output of the penultimate multi-head self-attention layer of our model is plotted, specifically for the positive class. Two types of CAM are shown: GradCAMElementWise47 and HiResCAM48. We have augmented the dataset in Supplementary Fig. 16 with seven additional experimental replicates to demonstrate the method’s performance.

Back to article page