A Calibrated Multiexit Neural Network for Detecting Urothelial Cancer Cells

Table 2

Results (in terms of test accuracy and test ECE) of all eight variants evaluated on the three datasets. TS refers to a posttraining temperature scaling operation, as described in Model Calibration.

			UC		B-BBCD		BBCD
Architecture	Loss	TS	Acc.	ECE	Acc.	ECE	Acc.	ECE
Standard	Cross-entropy	No	86.60%	4.77%	82.25%	3.57%	78.00%	5.83%
		Yes	86.60%	1.28%	82.25%	1.15%	78.00%	4.84%
	Focal	No	87.07%	1.50%	84.75%	3.11%	78.50%	3.72%
		Yes	87.07%	1.11%	84.75%	1.63%	78.50%	3.45%
Multiexit	Cross-entropy	No	89.13%	3.23%	85.09%	7.91%	84.00%	10.97%
		Yes	89.13%	2.30%	85.09%	2.78%	84.00%	5.21%
	Focal	No	89.90%	8.99%	87.07%	5.76%	85.04%	7.76%
		Yes	89.90%	1.43%	87.07%	1.57%	85.04%	1.97%