. 2024 Mar 18;14:6482. doi: 10.1038/s41598-024-56081-7

Table 2.

Five-fold testing: quantitative performance evaluation of the cell segmentation module (DL/IDL) compared to state-of-the-art methods.

		Detection metrics					Semantic metric	Speed metrics
		mIoU (%)	F1 (%)	Accuracy (%)	Precision (%)	Recall (%)	Dice (%)	Train epochs	Inference (s)
	Cellpose¹⁷	$63.85 \pm 1.39$	$83.35 \pm 1.41$	$72.22 \pm 1.99$	$91.30 \pm 1.75$	$77.09 \pm 1.24$	$84.51 \pm 1.16$	500	$1.032 \pm 0.0238$
	Stardist¹⁶	$66.21 \pm 0.96$	$85.82 \pm 0.93$	$76.04 \pm 1.31$	94.80±0.15	$78.95 \pm 1.47$	$87.86 \pm 0.52$	400	$0.222 \pm 0.0173$
Ours	Att-UNet+LSTM	$71.17 \pm 2.77$	$86.12 \pm 2.52$	$76.73 \pm 3.54$	$92.20 \pm 3.20$	$81.33 \pm 2.72$	$91.87 \pm 2.6$	40	$0.369 \pm 0.0062$
	Att-UNet (XAI)	73.55±1.41	$\underline{86.53 \pm 1.64}$	$\underline{77.00 \pm 2.47}$	$89.00 \pm 2.63$	84.53±1.55	$\underline{93.77 \pm 0.34}$	20	$\underline{0.157 \pm 0.0021}$
	UNet+LSTM	$72.23 \pm 1.95$	86.90±1.80	77.72±2.53	$\underline{93.47 \pm 1.61}$	$81.66 \pm 2.20$	94.04 ± 0.28	40	$0.369 \pm 0.0062$
	UNet (XAI)	$\underline{72.29 \pm 2.6}$	$85.44 \pm 2.12$	$75.72 \pm 2.92$	$87.95 \pm 2.87$	$\underline{83.56 \pm 2.46}$	$93.18 \pm 1.18$	20	0.136±0.0069
	BiONet+LSTM	$63.94 \pm 6.48$	$80.25 \pm 5.41$	$68.44 \pm 7.01$	$90.83 \pm 2.88$	$72.73 \pm 7.44$	$92.2 \pm 2.78$	40	$0.306 \pm 0.0062$
	BiONet (XAI)	$65.94 \pm 5.31$	$81.43 \pm 4.13$	$70.09 \pm 5.33$	$88.67 \pm 3.67$	$75.99 \pm 5.95$	$91.24 \pm 2.66$	20	$0.203 \pm 0.0042$

The reported results are the $(m e a n \pm s t a n d a r d d e v i a t i o n)$ , computed over 5-fold testing. The best metrics (per column) are highlighted in bold, and the second-best metrics are underlined. Instance-level segmentation (detection) evaluations were used to assess performance with different metrics (per cell mask). The mIoU (mean Intersection over Union) is calculated as the sum of IoU (cell mask-wise) of the predicted cell masks divided by the ground-truth cell count. To report these metrics, we used $I o U \geq 50 %$ between ground truth and predicted masks to compute TP, FP, and FN. The F1 score is defined as $F 1 = \frac{2 T P}{2 T P + F P + F N}$ , while the $A c c u r a c y = \frac{TP}{T P + F P + F N}$ , $P r e c i s i o n = \frac{TP}{T P + F P}$ , and $R e c a l l = \frac{TP}{T P + F N}$ . We utilized a semantic segmentation metric (i.e., Dice coefficient) to quantify the foreground/background pixel-wise separation, defined as $D i c e = \frac{2 | g t \cap p r e d |}{| g t | + | p r e d |}$ , where gt is the ground truth mask and pred is the predicted mask ( $b a c k g r o u n d = 0, f o r e g r o u n d = 1$ ). The training epochs refer to the number of epochs needed to complete the training phase. Inference time (on the test set) per image was computed using the following hardware: an 8-core i7 9700K CPU, 16GB RAM, NVIDIA MSI 2080.