Table 2.

Experiment 2: Training has been performed with samples from five slices and the remaining one was used for testing. We trained the forest with 198 samples per class and slice and no post-hoc smoothing was performed. The dash indicates that no labels are available for a certain slice/class combination. Average SE and PPV values are between 80 and 95 percent. We see that a Random Forest trained with samples from five out of the six slices generalizes reasonably well and can be used to classify the test slice.

		tissue class
slice	measure	necrotic	viable	gelatine	interface	glass/hole	mean
S3	SE	92.8	82.7	-	89.7	89.5	88.7
	PPV	57.6	96.8	-	41.9	98.6	73.7
S4	SE	64.7	98.3	-	89.9	97.1	87.4
	PPV	95.6	85.7	-	98.3	86.3	91.4
S5	SE	94.7	91.3	86.7	-	98.1	92.7
	PPV	86.7	96.8	100	-	99.4	95.8
S7	SE	99.4	75.4	99.1	99.6	89.1	92.5
	PPV	30.4	99.7	99.5	94.8	99.5	84.8
S9	SE	81.0	96.2	84.0	96.4	97.2	90.1
	PPV	54.0	97.3	99.4	60.0	97.0	81.5
S11	SE	96.4	90.4	87.4	99.1	91.0	92.8
	PPV	40.0	98.7	99.4	68.3	99.7	81.2