Skip to main content
Fig. 6 | BMC Bioinformatics

Fig. 6

From: Visualizing histopathologic deep learning classification and anomaly detection using nonlinear feature space dimensionality reduction

Fig. 6

Alternative dimensionality reduction approaches. a-b Hypothetical cartoon depicting advantages of the t-SNE plot as a “graded” and complementary classification tool. a Representative image spaces within a CNN as organized by different dimensionality reduction techniques. Unlike linear representations (e.g. PCA) with discrete and adjacent decision borders, the exaggerated separation of classes on the t-SNE plot provides more distinct categorical classes for testing. This key difference has significant advantages for evaluating the distribution of test tiles amongst previously trained classes. As shown, this could allow more effective handling of variants of already trained cases (“blue square variant”, Panel a) and true “undefined” classes (green square, Panel b). c-d For comparison, we show the PCA depiction of our trained CNN. Although similar in the overall arrangement of classes, there is notably little separation between tissue types. This difference leads to the vast majority (71%) of cases being classified as “differential” and a substantial decrease in performance (13% correct and 16.2% errors compared to both t-SNE and probability distribution scoring (Prob). Correct cases of the “combined” analysis represents those where both the t-SNE and Prob score were in agreement. The undefined class in the “combined” analysis represents cases with classification discordance between the two methods

Back to article page