Sinus Detection — Model Comparison

Sinus Detection for Maxillary Sinus Segmentation (Paper Layout)

Metric	YOLOv11n	YOLOv11n-aug	YOLOv11s	YOLOv11s-aug	YOLOv8n	YOLOv8n-aug	YOLOv8s	YOLOv8s-aug
Detection Rate	1.0000	1.0000	1.0000	1.0000	0.8485	1.0000	1.0000	1.0000
IoU Mean ± Std	0.8050 ± 0.0731	0.8038 ± 0.0814	0.7989 ± 0.1186	0.8201 ± 0.0752	0.6677 ± 0.3442	0.8516 ± 0.0778	0.7887 ± 0.0866	0.8138 ± 0.0699
DSC Mean ± Std	0.8900 ± 0.0481	0.8888 ± 0.0550	0.8809 ± 0.1135	0.8991 ± 0.0499	0.7274 ± 0.3572	0.9178 ± 0.0500	0.8791 ± 0.0575	0.8956 ± 0.0457
Centroid Dist (px)	7.89 ± 3.99	9.49 ± 4.28	11.11 ± 13.26	8.58 ± 3.10	41.92 ± 64.71	7.83 ± 4.89	10.74 ± 7.11	8.77 ± 3.37
95% CI IoU	[0.7865, 0.8220]	[0.7828, 0.8221]	[0.7653, 0.8228]	[0.8005, 0.8364]	[0.5780, 0.7561]	[0.8309, 0.8687]	[0.7661, 0.8088]	[0.7953, 0.8292]
95% CI DSC	[0.8777, 0.9009]	[0.8742, 0.9009]	[0.8485, 0.9013]	[0.8859, 0.9096]	[0.6313, 0.8179]	[0.9042, 0.9284]	[0.8640, 0.8923]	[0.8833, 0.9055]

Metric	YOLOv11n	YOLOv11s	YOLOv8n	YOLOv8s
Detection Rate	1.0000	1.0000	0.8485	1.0000
IoU Mean ± Std	0.8050 ± 0.0731	0.7989 ± 0.1186	0.6677 ± 0.3442	0.7887 ± 0.0866
DSC Mean ± Std	0.8900 ± 0.0481	0.8809 ± 0.1135	0.7274 ± 0.3572	0.8791 ± 0.0575
Centroid Dist (px)	7.89 ± 3.99	11.11 ± 13.26	41.92 ± 64.71	10.74 ± 7.11
95% CI IoU	[0.7865, 0.8220]	[0.7653, 0.8228]	[0.5780, 0.7561]	[0.7661, 0.8088]
95% CI DSC	[0.8777, 0.9009]	[0.8485, 0.9013]	[0.6313, 0.8179]	[0.8640, 0.8923]

Metric	YOLOv11n-aug	YOLOv11s-aug	YOLOv8n-aug	YOLOv8s-aug
Detection Rate	1.0000	1.0000	1.0000	1.0000
IoU Mean ± Std	0.8038 ± 0.0814	0.8201 ± 0.0752	0.8516 ± 0.0778	0.8138 ± 0.0699
DSC Mean ± Std	0.8888 ± 0.0550	0.8991 ± 0.0499	0.9178 ± 0.0500	0.8956 ± 0.0457
Centroid Dist (px)	9.49 ± 4.28	8.58 ± 3.10	7.83 ± 4.89	8.77 ± 3.37
95% CI IoU	[0.7828, 0.8221]	[0.8005, 0.8364]	[0.8309, 0.8687]	[0.7953, 0.8292]
95% CI DSC	[0.8742, 0.9009]	[0.8859, 0.9096]	[0.9042, 0.9284]	[0.8833, 0.9055]

Metric	YOLOv11n	YOLOv11n-aug	YOLOv11s	YOLOv11s-aug	YOLOv8n	YOLOv8n-aug	YOLOv8s	YOLOv8s-aug
Precision	0.995	0.995	1.000	0.958	0.895	0.967	0.994	0.990
Recall	0.915	0.941	0.934	0.941	0.735	0.971	0.912	0.912
mAP50	0.955	0.946	0.972	0.955	0.938	0.976	0.977	0.950
mAP50-95	0.687	0.672	0.683	0.694	0.624	0.651	0.605	0.724

Metric	YOLOv11n	YOLOv11s	YOLOv8n	YOLOv8s
Precision	0.995	1.000	0.895	0.994
Recall	0.915	0.934	0.735	0.912
mAP50	0.955	0.972	0.938	0.977
mAP50-95	0.687	0.683	0.624	0.605

Metric	YOLOv11n-aug	YOLOv11s-aug	YOLOv8n-aug	YOLOv8s-aug
Precision	0.995	0.958	0.967	0.990
Recall	0.941	0.941	0.971	0.912
mAP50	0.946	0.955	0.976	0.950
mAP50-95	0.672	0.694	0.651	0.724

Figure 1. Segmentation metrics comparison across 4 architectures with offline augmentation (higher is better).

Figure 2. Segmentation metrics comparison across 4 architectures without offline augmentation (higher is better).

Figure 3a. Validation metrics — with augmentation.

Figure 3b. Validation metrics — without augmentation.

Figure 4a. Radar plot — test + validation metrics with augmentation.

Figure 4b. Radar plot — test + validation metrics without augmentation.