Table 2.

Comparison of performance between different models for predicting 4-category molecular subtypes of breast cancer in the test cohort (n = 672).

Method	Modality	Accuracy (%)	Precision (%)	Recall (%)	F1-score	MCC
US-ResNet50	US	81.1 [78.0, 83.9]	79.7 [76.3, 82.9]	76.9 [73.3, 80.2]	0.774 [0.739, 0.807]	0.731 [0.689, 0.772]
MG-ResNet50	MG	82.0 [79.0, 84.8]	81.3 [77.8, 84.5]	78.3 [74.9, 81.5]	0.787 [0.753, 0.821]	0.744 [0.704, 0.785]
Multi-ResNet50	MG + US	84.4 [81.5, 87.1]	83.7 [80.4, 86.9]	81.3 [77.8, 84.5]	0.820 [0.786, 0.852]	0.777 [0.736, 0.814]
Multi-ResNet34	MG + US	83.3 [80.4, 86.0]	83.1 [80.0, 86.3]	79.8 [76.5, 83.1]	0.803 [0.769, 0.838]	0.764 [0.724, 0.803]
Multi-ResNet101	MG + US	83.0 [80.1, 85.9]	82.5 [79.1, 85.7]	79.4 [76.1, 82.6]	0.799 [0.766, 0.832]	0.759 [0.720, 0.798]
Multi-Inceptionv3	MG + US	82.6 [79.6, 85.4]	82.0 [78.7, 85.2]	78.9 [75.7, 82.0]	0.794 [0.761, 0.827]	0.753 [0.713, 0.793]
Multi-ResNet34+SE	MG + US	85.1 [82.3, 87.8]	85.1 [82.0, 88.0]	82.1 [78.8, 85.2]	0.826 [0.794, 0.857]	0.790 [0.751, 0.826]
Multi-ResNet50+SE	MG + US	86.0 [83.3, 88.5]	86.0 [82.9, 88.8]	82.5 [79.2, 85.7]	0.835 [0.802, 0.867]	0.801 [0.763, 0.838]
Multi-ResNet101+SE	MG + US	84.8 [82.1, 87.5]	84.4 [81.5, 87.2]	81.6 [78.5, 84.8]	0.822 [0.791, 0.854]	0.784 [0.749, 0.822]
Multi-Inceptionv3+SE	MG + US	83.3 [80.4, 86.2]	82.3 [79.1, 85.6]	79.7 [76.3, 82.9]	0.800 [0.767, 0.833]	0.764 [0.724, 0.802]
Proposed (MDL-IIA)	MG + US	88.5 [86.0, 90.9]	87.8 [85.0, 90.7]	85.4 [82.2, 88.4]	0.862 [0.831, 0.892]	0.837 [0.803, 0.870]

Values in brackets are 95% confidence intervals [95% CI, %].

MG mammography, US ultrasound, SE Squeeze-and-Excitation, MCC Matthews correlation coefficient, MDL-IIA multi-modal deep learning with intra- and inter-modality attention modules.