Results Based on Consensus Manual Classification

Table 3.5: Confusion matrices for MIAS classification according to BIRADS categories using the consensus classification. The results are based on a leave-one-woman-out methodology with $322$ mammograms. (a) kNN classifier, (b) C$4.5$ decision tree, and (c) Bayesian classifier.

               kNN ( $77\%,\kappa = 0.68$ )

C4.5 ( $72\%,\kappa = 0.61$ )

Bayesian ( $86\%,\kappa = 0.81$ )

		B-I	B-II	B-III	B-IV
	B-I	$70$	$13$	$1$	$3$
		B-II	$9$	$80$	$13$
		B-III	$1$	$17$	$73$
		B-IV	$3$	$2$	$8$

B-I	B-II	B-III	B-IV
$72$	$13$	$1$	$1$
$13$	$68$	$20$	$2$
0	$21$	$68$	$6$
0	$2$	$11$	$24$

B-I	B-II	B-III	B-IV
$79$	$1$	$3$	$4$
$3$	$86$	$6$	$8$
0	$2$	$85$	$8$
0	$6$	$4$	$27$

(a)

(b)

(c)

Table  shows results based on a leave-one-woman-out methodology for the classification of the whole MIAS database according to the consensus ground truth. The performance of the individual classifiers is $77\%$ correct classification for kNN and $72\%$ for C$4.5$ . These are intermediate values between Expert A and both Expert B and C. However, the Bayesian combination of the classifiers results in improvement and $86\%$ correct classification is achieved, which gives a better performance compared to those obtained by the individual experts without consensus. This result is confirmed by $\kappa=0.81$ , which belongs to the Almost Perfect category. Examining each class alone, BIRADS I reached $91\%$ correct classification, BIRADS II $84\%$ , BIRADS III $89\%$ , and BIRADS IV $73\%$ .

Using the low/high density division, low density mammograms are $89\%$ correctly classified, while high density ones reach $94\%$ , resulting in an overall two class classification equal to $91\%$ .