Table 5.

Diagnostic performance comparison for both classification stages between the developed RC-CAD system and other classification approaches (e.g., random forest (RF) and support vector machine (SVM)). Using leave-one-subject-out (LOSO) and a randomly stratified 10-fold cross-validation approach, the diagnostic abilities of the RC-CAD outperformed the others. Let Sens: sensitivity, Spec: Specificity, DSC: Dice similarity coefficient, and Acc: Accuracy.

First Stage Classification (RCC vs. AML) Performance (Mean ± SD ≈)
Method	Validation	Sens%	Spec%	DSC
RC-CAD (Proposed)	LOSO	95.3 ± 2.0	99.9 ± 0.4	0.98 ± 0.01
	10-fold	89.0 ± 3.4	91.0 ± 2.7	0.90 ± 0.02
RFs	LOSO	89.0 ± 1.7	92.7 ± 2.7	0.91 ± 0.02
	10-fold	88.4 ± 1.0	90.7 ± 3.0	0.89 ± 0.01
SVM${}_{Quad}$	LOSO	82.9 ± 0.0	88.6 ± 0.0	0.85 ± 0.00
	10-fold	81.9 ± 2.2	87.7 ± 2.5	0.84 ± 0.02
Second Stage Classification (ccRCC vs. nccRCC) Performance (Mean ± SD ≈)
Method	Validation		Acc%
RC-CAD (Proposed)	LOSO		89.6 ± 5.0
	10-fold		78.6 ± 5.7
RFs	LOSO		53.7 ± 3.7
	10-fold		51.9 ± 2.6
SVM${}_{Quad}$	LOSO		52.9 ± 0.0
	10-fold		54.3 ± 3.0

Hyperparameters: MLP-ANN (optimization function: trainlm, max epochs = 500, hidden layers: hl${}_1$ = 50 nodes, hl${}_2$ = 25 nodes, goal = 0, max validation failure = 6, min gradient = ${10}^{-7}$, training gain ($\mu$): initial $\mu$ = 0.001, $\mu$ decrease factor = 0.1, $\mu$ increase factor = 10, max $\mu$ = 1e${}^{10}$); RF (method: Bag, number of learning cycles = 30); SVM (kernel function: quadratic, box constraint = 1).