Skip to main content
. 2021 Mar 22;11:6482. doi: 10.1038/s41598-021-85758-6

Table 6.

Using variance in predicted tumor likelihood amongst the three NN configurations (inter-NN variance, VER) to identify tumors predicted by at least one but not all NNs.

Figure 6 Histologya Regionb NN tumor prediction Inter-NN variance (VRA) Intra-NN variance (VER)
FP HW FPHW FP HW FPHW
s1 100 B 0.615 0.871 0.785 0.017 0.097 0.006 0.019
s2 100 T 0.635 0.927 0.846 0.023 0.097 0.002 0.008
s3 – T 0.606 0.932 0.800 0.027 0.13 0.001 0.031
s4 100 B 0.448 0.897 0.777 0.054 0.103 0.002 0.039
s5 100 B 0.366 0.891 0.674 0.07 0.062 0.002 0.06
s6 100 B 0.354 0.873 0.487 0.072 0.066 0.010 0.088
s7 75 T 0.141 0.763 0.387 0.098 0.003 0.047 0.07

Significant inter-NN variance (see Fig. 5B, VER > 1σ) occurred in 7 targets, all from tumor or boundary regions. The highest NN tumor probability for each target appears in bold. NNs using only FP inputs predicts 3 of these targets contain tumor, but with relatively low probabilities (PTumor = 0.62 ± 0.02, N = 3). FPHW using the full nine inputs predicts 5 of the 7 targets contain tumor and produces higher probabilities (PTumor = 0.78 ± 0.06, N = 5). The NNs trained only on HW data predict all 7 of these targets contain tumor. HW NNs generated the highest tumor likelihood probabilities (PTumor = 0.88 ± 0.06, N = 3). The variation in the output of each stochastic NN (train-test cycle repeated 10 times with random weight restarts) provides an intra-NN variance or VRA for each of the three NN configurations. Six of the seven HW NNs exhibits the least variance in their predictions (VRA between 0.001 and 0.10). One a single target (s7) NN FP exhibits minimal variance across ten trials (VRA ~ 0.003) while generating a low tumor probability (P ~ 0.141). For target s7 HW predicts a tumor likelihood of P ~ 0.763 with VRA = 0.047.

aPathologist target labels in Fig. 6. Pathologist estimate of likelihood probe would strike tumor tissue in 1 mm2 region around laser.

bMacroscopic (1X) visual assessment of spectra collection sites as tumor (T), healthy (H), or boundary (B) regions.