Improving Accuracy of Lung Nodule Classification Using Deep Learning with Focal Loss

Table 2

Performance of LdcNet for different architectures and hyperparameter sets.

#	Input and network			Focal loss		Performance (%)
	Input scale	Batch size	# conv. blocks	γ	α	Accuracy	Sensitivity	Specificity
1	0.688	128	2	1	1	92.7	77.7	96.2
2	0.594	160	2	1	0.5	94.3	80.3	93.5
3	0.906	256	2	2.5	0.5	94.5	80.1	95.1
4	1.0	224	2	1.5	0.2	95.3	86.7	95.2
5	0.844	224	3	2.5	0.2	97.0	82.7	98.2
6	0.875	256	3	2	0.5	97.2	94.7	97.4
7	0.906	224	3	2.5	0.5	97.2	96.0	97.3
8	0.938	224	3	1.5	0.5	97.3	93.3	97.6
9	0.844	256	4	1.5	0.5	96.8	93.7	97.4
10	0.938	160	4	2	0.7	97.0	85.0	98.1
11	0.844	256	4	1.5	1	97.1	91.3	97.7
12	0.875	128	4	1.5	0.5	97.1	92.0	97.5