Diabetic Retinopathy Lesion Segmentation Based on Hierarchical Feature Progressive Fusion in Retinal Fundus Images

Pengchao Ding; Feng Li

doi:10.3788/CJL240731

Chinese Journal of Lasers, Volume. 51, Issue 21, 2107107(2024)

Diabetic Retinopathy Lesion Segmentation Based on Hierarchical Feature Progressive Fusion in Retinal Fundus Images

Pengchao Ding and Feng Li^*

School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Overall architecture of the proposed PMFF-Net model

Download full size

Fig. 2. Hybrid Transformer module

Download full size

Fig. 3. Gradual characteristic fusion module

Download full size

Fig. 4. Dynamic attention module

Download full size

Fig. 5. Selective edge aggregation module

Download full size

Fig. 6. Visualization of segmentation results for each model in the IDRiD dataset (red: EX; green: HE; blue: MA; pink: SE)

Download full size

Fig. 7. Visualization of segmentation results for each model in the DDR dataset (red: EX; green: HE; blue: MA; pink: SE)

Download full size

Fig. 8. Visualization of generalizability experimental results for each model on the DDR dataset (red: EX; green: HE; blue: MA; pink: SE)

Download full size

Fig. 9. Visualization of generalizability experimental results for each model on the IDRiD dataset (red: EX; green: HE; blue: MA; pink: SE)

Download full size

Fig. 10. Ablation experiment visualization results (red: EX; green: HE; blue: MA; pink: SE)

Download full size

Table 1. Comparison of experimental results for different models on the IDRiD dataset

View table

Table 1. Comparison of experimental results for different models on the IDRiD dataset

Model	Year	Dice /%				mDice /%	mIoU /%	Time /ms
Model	Year	EX	HE	MA	SE	mDice /%	mIoU /%	Time /ms
U-Net^［14］	2015	36.49	41.88	13.03	53.12	36.13	26.10	28.87
Dense U-Net^［16］	2018	47.53	48.32	18.54	44.57	39.74	28.37	29.49
DeepLabv3+^［13］	2018	48.19	49.21	18.29	46.93	40.66	28.91	29.62
U-Net++^［15］	2018	39.34	33.06	18.42	40.73	32.89	22.40	31.30
PMCNet^［17］	2022	46.84	49.47	17.47	48.45	40.56	28.70	29.02
H2Former^［25］	2023	49.91	47.42	16.99	50.37	41.17	30.11	33.82
DS2F^［31］	2024	47.74	47.83	19.40	59.17	43.53	32.33	32.86
PMFF-Net	2024	52.00	50.83	18.18	59.43	45.11	33.39	34.74

Table 2. Comparison of experimental results for different models on the DDR dataset

View table

Table 2. Comparison of experimental results for different models on the DDR dataset

Model	Year	Dice /%				mDice /%	mIoU /%	Time /ms
Model	Year	EX	HE	MA	SE	mDice /%	mIoU /%	Time /ms
U-Net^［14］	2015	11.83	18.47	13.73	50.99	23.76	22.26	30.96
Dense U-Net^［16］	2018	24.73	17.10	19.94	45.67	26.86	24.39	32.74
DeepLabv3+^［13］	2018	37.90	23.90	21.57	40.13	30.88	28.40	30.28
U-Net++^［15］	2018	14.58	9.62	17.76	42.92	21.22	19.81	35.12
PMCNet^［17］	2022	38.38	25.28	17.08	40.37	30.28	27.61	31.36
H2Former^［25］	2023	39.94	22.68	15.00	47.28	31.47	28.89	37.84
DS2F^［31］	2024	40.54	24.55	18.21	54.57	34.47	31.89	35.66
PMFF-Net	2024	40.78	29.27	21.39	55.11	36.64	35.04	38.48

Table 3. Comparison of generalizability of different models on DDR dataset (training and testing on IDRiD and DDR datasets, respectively)

View table

Table 3. Comparison of generalizability of different models on DDR dataset (training and testing on IDRiD and DDR datasets, respectively)

Model	Year	Dice /%				mDice /%	mIoU /%
Model	Year	EX	HE	MA	SE	mDice /%	mIoU /%
U-Net^［14］	2015	13.62	5.94	4.99	26.94	12.88	11.97
Dense U-Net^［16］	2018	14.50	19.52	13.95	22.47	17.61	16.46
DeepLabv3+^［13］	2018	16.32	17.94	17.10	30.64	20.50	18.85
U-Net++^［15］	2018	8.37	4.80	2.52	25.62	10.33	9.44
PMCNet^［17］	2022	27.66	7.45	15.32	29.97	20.10	18.72
H2Former^［25］	2023	26.32	14.90	13.19	30.55	21.24	20.03
DS2F^［31］	2024	27.88	8.77	24.32	32.18	23.29	22.15
PMFF-Net	2024	16.45	11.44	30.77	44.66	25.83	23.59

Table 4. Comparison of generalizability of different models on the IDRiD dataset (training and testing on DDR and IDRiD datasets, respectively)

View table

Table 4. Comparison of generalizability of different models on the IDRiD dataset (training and testing on DDR and IDRiD datasets, respectively)

Model	Year	Dice /%				mDice /%	mIoU /%
Model	Year	EX	HE	MA	SE	mDice /%	mIoU /%
U-Net^［14］	2015	26.95	18.40	8.03	14.28	16.92	10.05
Dense U-Net^［16］	2018	30.15	17.89	9.97	27.01	21.25	13.59
DeepLabv3+^［13］	2018	33.64	20.26	10.76	38.99	25.91	14.33
U-Net++^［15］	2018	25.60	16.70	6.30	13.74	15.59	9.50
PMCNet^［17］	2022	38.62	18.49	11.71	31.24	25.01	16.01
H2Former^［25］	2023	37.78	19.96	10.68	37.69	26.53	16.94
DS2F^［31］	2024	40.94	23.39	11.74	42.64	29.68	20.15
PMFF-Net	2024	35.57	30.66	12.19	49.11	31.88	22.17

Table 5. Comparison of ablation experimental results

View table

Table 5. Comparison of ablation experimental results

Module	Dice /%				mDice /%	mIoU /%
Module	EX	HE	MA	SE	mDice /%	mIoU /%
Baseline	47.26	50.03	16.68	48.64	40.65	29.77
Baseline+HT	47.91	50.42	16.99	49.37	41.17	30.11
Baseline+HT+SEA	51.33	51.14	18.34	52.54	43.34	31.83
Baseline+HT+SEA+GCF	51.95	51.70	15.52	57.56	44.18	32.79
Baseline+HT +SEA+GCF+DA	52.00	50.83	18.18	59.43	45.11	33.39

Tools

Get Citation

Copy Citation Text

Pengchao Ding, Feng Li. Diabetic Retinopathy Lesion Segmentation Based on Hierarchical Feature Progressive Fusion in Retinal Fundus Images[J]. Chinese Journal of Lasers, 2024, 51(21): 2107107

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Biomedical Optical Imaging

Received: Apr. 1, 2024

Accepted: May. 21, 2024

Published Online: Oct. 31, 2024

The Author Email: Feng Li (lifenggold@163.com)

DOI:10.3788/CJL240731

CSTR:32183.14.CJL240731

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology