Image detection method for multi-category lesions in wireless capsule endoscopy based on deep learning models

Table 1 Experimental parameter settings

Parameter name	Parameter value
Initial learning rate	0.01
Learning rate float	0.01
Epochs	300
Batch size	16
Optimizer	AdamW
Weight_decay	0.0005
Momentum	0.937

Table 2 Training results of the YOLOv8 versions using our dataset

Model	Parameter (M)	mAP50	mAP50:95	GFLOPS	FPS
YOLOv8n	2.87	86.0	66.2	8.2	400.00
YOLOv8s	10.62	86.1	67.6	28.7	283.03
YOLOv8m	24.67	88.6	68.3	79.1	189.47
YOLOv8 L	41.62	85.3	67.1	165.5	149.25
YOLOv8x	65.01	83.2	64.1	258.2	114.94

Table 3 Training results of different detection models on the dataset

Model	mAP50	mAP50:90	FPS
YOLOv5	78.3	58.9	176.25
YOLOv6	79.6	58.1	316.63
YOLOv7	75.8	56.4	206.37
YOLOv9[32]	82.0	64.7	94.30
SSD	78.6	57.9	84.62
Faster R-CNN	79.0	60.9	61.98
RT-DETR[33]	81.9	63.9	143.46
WCE_detection	91.5	68.6	129.70

WCE: Wireless capsule endoscopy.

Table 4 Experimental results of different numbers of detection heads

Method	P2	P3	P4	P5	mAP50	GFLOPS	FPS
Method 1	Y	N	N	N	87.1	47.9	233.65
Method 2	N	Y	N	N	88.6	79.1	189.47
Method 3	N	N	Y	N	89.2	126.7	179.19
Method 4	N	N	N	Y	89.5	166.3	153.95

Y: Means that the module is designed on the original YOLOv8 algorithm for the experiment; N: Means that the module was not used in the model design.

Table 5 Experimental results of different necks

Method	FPN	PANet	BiFPN	mAP50	GFLOPS	FPS
Method 1	Y	N	N	86.3	58.6	226.63
Method 2	N	Y	N	88.6	79.1	189.47
Method 3	N	N	Y	90.1	98.4	178.36

Y: Means that the module is designed on the original YOLOv8 algorithm for the experiment; N: Means that the module was not used in the model design.

Table 6 Experimental results of different attention modules

Method	CBAM	SE	VIT	Swin Transformer	mAP50	GFLOPS	FPS
Method 1	Y	N	N	N	88.9	133.6	163.96
Method 2	N	Y	N	N	87.2	121.5	173.73
Method 3	N	N	Y	N	89.2	219.9	103.76
Method 4	N	N	N	Y	90.7	159.7	149.58

Y: Means that the module is designed on the original YOLOv8 algorithm for the experiment; N: Means that the module was not used in the model design.

Table 7 Ablation experiment results

Method	p4	BiFPN	Swin Transformer	mAP50	GFLOPS	FPS
Method 1	N	N	N	88.6	79.1	189.47
Method 2	Y	N	N	89.2	126.7	179.19
Method 3	Y	Y	N	90.6	138.4	183.35
Method 4	Y	Y	Y	91.5	203.6	129.70

Y: Means that the module is designed on the original YOLOv8 algorithm for the experiment; N: Means that the module was not used in the model design.