Classification Performance Based on Human Perceptual Features

Wu et al. compared the accuracy of radiologists to that of a computerized technique based on 14 features perceived by radiologists and an ANN classifier [54]. The database included 30 mammograms with masses and 30 with microcalcification clusters, and the biopsy results indicated that 26 mammograms were malignant while the rest were benign. The diagnostic performance of radiologists was obtained from reports of five attending and five resident radiologists experienced in reading mammograms. The accuracy levels of computer classification and attending and resident radiologists, measured with A' , were 0.89, 0.84, and 0.80, respectively, with a statistically significant difference between computer and both the attending and resident radiologists (p < 0.01).

Similar results were obtained in a study that used 18 features based on BI-RADS and an ANN applied to 206 mammograms where 73 were shown to be malignant in biopsy [55]. The AZ accuracy levels in this study were 0.89 for the computer and 0.85 for the radiologists. Although these two A' values were not statistically different (p = 0.29) the differences in specificity observed at high sensitivity levels were statistically significant [55].

0 0

Post a comment