Figure 3

Optimization and performance of the nearest-neighbor classifier for diagnostic discrimination. (a) Nearest-neighbor number optimization by using the receiver operating characteristic curve for discrimination between all pathology subtypes. (b) A confusion matrix showing the true and predicted diagnosis for all spectroscopic measures according to the nearest-neighbor classifier, presented as a percentage of the total number of diagnosed pixels (N). A multiparametric diagnostic map for a malignant (row 1) and benign (row 2) tissue is illustrated (c through e); the patient histology is shown in column (c), a photograph of the tissue imaged by SFDI is shown in column (d), and the patient-specific diagnostic map generated by the classifier is shown in column (e).