Multi-variable heatmaps for computer-aided diagnostic models

Abstract

Introduced here are diagnostic platforms able to produce visualizations that visually highlight the pixels considered as evidence of a medical condition by a computer-aided diagnostic (CADx) model. A CADx model may be based on, for example, a multi-headed neural network designed to detect the presence/progression of multiple medical conditions. By explaining how each output is produced by the multi-headed neural network, a diagnostic platform can identify the latent variable (s) responsible for producing each output. For example, a diagnostic platform may create a multi-variable heatmap that visually distinguishes the pixels considered as evidence of each multiple condition by a multi-headed neural network. To accomplish this, the diagnostic platform may create multiple heatmaps by producing, for each output, a separate heatmap that distinguishes the pixels in the image considered as evidence of the corresponding medical condition by the multi-headed neural network, and then compiling the multiple heatmaps into the multi-variable heatmap.