TABLE IV. Single- and Multi-Time-Point Networks on NLSTx.

	Single-time-point NLSTx
		Interval t0			Interval t1			Interval t2			All Intervals
Network	Params	ROC-AUC	AP	NLL	ROC-AUC	AP	NLL	ROC-AUC	AP	NLL	ROC-AUC	AP	NLL
CNN-3D [12]	44.4M	0.798	0.538	0.472	0.815	0.666	0.476	0.841	0.585	0.418	0.817	0.597	0.456
CRN1-R [12]	16.4M	0.737	0.461	0.529	0.765	0.548	0.535	0.774	0.496	0.512	0.759	0.502	0.526
CRN1-2D [12]	16.4M	0.806	0.559	0.396	0.835	0.669	0.407	0.855	0.628	0.337	0.831	0.620	0.381
CAN1-R	14.9M	0.722	0.435	0.580	0.764	0.557	0.568	0.755	0.474	0.565	0.747	0.488	0.571
CAN1-2D	14.9M	0.826	0.641	0.385	0.868	0.741	0.371	0.879	0.690	0.330	0.858	0.691	0.362
	Multi-time-point NLSTx
		Interval t0-t1			Interval t1-t2			Interval t0-t1-t2
Network	Params	ROC-AUC	AP	NLL	ROC-AUC	AP	NLL	ROC-AUC	AP	NLL
CRN2-2D [12]	16.4M	0.847	0.684	0.348	0.849	0.659	0.331	-	-	-
CAN2-2D	15.7M	0.879	0.773	0.305	0.884	0.757	0.322	-	-	-
CRN3-2D [12]	16.4M	-	-	-	-	-	-	0.823	0.629	0.396
CAN3-2D	18.4M	-	-	-	-	-	-	0.882	0.745	0.328

Comparison of networks on the NLSTx dataset when tested on single- and multi-time-points (results from proposed networks for single- and multi-time-point classification in bold). Clear improvements in performance are seen by the proposed approach, CAN3-2D, for single-time-point classification when using ImageNet weights and the attention mechanism. Furthermore, the proposed approach uses less than half the network parameters than the closest equivalent 3-D network, CNN-3D. For multi-time-point classification, further gains are seen by using a one- and two-branch attention-based networks with inputs at multiple time points. Please see text for description of metrics.