ROC Analysis

ROC analysis for the evaluation of mass detection algorithms is closely related to ROC analysis for classification evaluation. Thus, each pixel of the image is treated as an instance of the classification process, and thus it can be a pixel belonging to a mass or a pixel not belonging to the mass. Therefore, this pixel is compared to the same pixel on the image obtained from the manually segmented image, resulting then in a well-classified pixel or a bad-classified pixel. A pixel is well-classified if in both images its state is the same: mass (in this case, it is a true positive) or not mass (true negative). A pixel is bad-classified if it has different state in both images: if the pixel is classified as not mass by the radiologist and classified as mass in the CAD system (false positive) or the inverse case (false negative). Table  summarizes this notation. Note that is closely related to Table .

Table C.4: Description of the notation used in CAD diagnosis. TP and TN represents the pixels classified correctly by the CAD systems. FP and FN are the bad classified pixels

		Definition
		Ground-Truth	Segmentation
	True Positive (TP)	Mass	Mass
		True Negative (TN)	No Mass
		False Positive (FP)	No Mass
		False Negative (FN)	Mass