Laser & Optoelectronics Progress, Volume. 62, Issue 4, 0415003(2025)
Polyp Segmentation Based on GODC-U-Net Model
Figures & Tables(15)
Fig. 5. Partial experimental results of polyp image segmentation based on GODC-U-Net. (a)‒(d) Original input images; (e)‒(h) ground truth masks; (j)‒(m) model predicted images
Table 1. Polyp segmentation datasets used in the experiment
View tableTable 1. Polyp segmentation datasets used in the experiment
Dataset Resolution /（pixel×pixel） Number of items Kvasir-SEG［ 16 ］720×576‒1920×1072 500 CVC-ClinicDB［ 17 ］384×288 612 CVC-ColonDB［ 18 ］500×288 300 ETIS-LaribPolyDB［ 19 ］1255×966 196
Table 2. Comparison of loss function methods
View tableTable 2. Comparison of loss function methods
Loss function mDice mIoU 0.5 Dice loss + 0.5 Focal loss 0.9246 0.8519 0.3 Dice loss + 0.7 Focal loss 0.9234 0.8512 0.7 Dice loss + 0.3 Focal loss 0.9293 0.8529
Table 3. Performance of different optimizers
View tableTable 3. Performance of different optimizers
Optimizer mDice mIoU Adam 0.9183 0.8439 SGDW 0.9067 0.8426 DiffGrad 0.9031 0.8408 AdamP 0.9254 0.8512 AdamW 0.9289 0.8527
Table 4. Comparison of different models on the Kvasir-SEG dataset
View tableTable 4. Comparison of different models on the Kvasir-SEG dataset
Method mDice mIoU Precision Recall Accuracy U-Net［ 13 ］0.8655 0.7629 0.8593 0.8718 0.9563 HarDNet-DFUS［ 25 ］0.8626 0.7584 0.9351 0.8005 0.9583 HRNetV2［ 26 ］0.8536 0.7438 0.8778 0.8297 0.9539 MSRF-Net［ 10 ］0.8508 0.7404 0.8993 0.8074 0.9543 Polyp2Seg［ 27 ］0.9183 0.8828 0.9274 0.8537 0.9713 PEFNet［ 28 ］0.8921 0.8336 0.9031 0.8354 0.9612 MSNet［ 29 ］0.9071 0.8622 0.9264 0.9037 0.9689 PraNet［ 8 ］0.9094 0.8339 0.9599 0.8647 0.9738 Polyp-Mixer［ 9 ］0.9167 0.8641 0.9383 0.8975 0.9653 Proposed 0.9293 0.8729 0.9185 0.9228 0.9786
Table 5. Comparison of different models on the CVC-ClinicDB dataset
View tableTable 5. Comparison of different models on the CVC-ClinicDB dataset
Method mDice mIoU Precision Recall Accuracy U-Net［ 13 ］0.7623 0.6171 0.7989 0.7303 0.9599 HarDNet-DFUS［ 25 ］0.7036 0.5682 0.8748 0.5985 0.9648 HRNetV2［ 26 ］0.7776 0.6361 0.826 0.7346 0.9629 MSRF-Net［ 10 ］0.9067 0.8282 0.9547 0.8621 0.9842 Polyp2Seg［ 27 ］0.7794 0.6298 0.8294 0.7464 0.9679 PEFNet［ 28 ］0.8665 0.8144 0.9486 0.8584 0.9816 MSNet［ 29 ］0.9216 0.8792 0.9803 0.9086 0.9891 PraNet［ 8 ］0.8742 0.7766 0.9608 0.8028 0.9786 Polyp-Mixer［ 9 ］0.9384 0.8936 0.9713 0.8365 0.9784 Proposed 0.9335 0.9063 0.9787 0.9165 0.9919
Table 6. Comparison of different models on the CVC-ColonDB dataset
View tableTable 6. Comparison of different models on the CVC-ColonDB dataset
Method mDice mIoU Precision Recall Accuracy U-Net［ 13 ］0.8032 0.7037 0.8164 0.8274 0.9807 HarDNet-DFUS［ 25 ］0.7398 0.5874 0.9572 0.6057 0.9761 HRNetV2［ 26 ］0.6383 0.4687 0.5858 0.701 0.9565 MSRF-Net［ 10 ］0.8371 0.7198 0.8603 0.8151 0.9829 Polyp2Seg［ 27 ］0.8086 0.7273 0.8317 0.8481 0.9812 PEFNet［ 28 ］0.7186 0.6381 0.8159 0.8292 0.9786 MSNet［ 29 ］0.7557 0.6785 0.8472 0.8673 0.9874 PraNet［ 8 ］0.9131 0.8401 0.9657 0.8659 0.9901 Polyp-Mixer［ 9 ］0.7916 0.7061 0.8438 0.8591 0.9837 Proposed 0.9142 0.8734 0.9556 0.9536 0.9936
Table 7. Comparison of different models on the ETIS-LaribPolypDB dataset
View tableTable 7. Comparison of different models on the ETIS-LaribPolypDB dataset
Method mDice mIoU Precision Recall Accuracy U-Net［ 13 ］0.7864 0.6794 0.7867 0.8026 0.9786 HarDNet-DFUS［ 25 ］0.8467 0.7197 0.9017 0.7619 0.9824 HRNetV2［ 26 ］0.5079 0.3495 0.5037 0.5194 0.9531 MSRF-Net［ 10 ］0.8067 0.6729 0.9432 0.7153 0.9824 Polyp2Seg［ 27 ］0.8825 0.7353 0.7618 0.8734 0.9373 PEFNet［ 28 ］0.8364 0.6973 0.8334 0.7315 0.9863 MSNet［ 29 ］0.7191 0.6647 0.7729 0.6967 0.9646 PraNet［ 8 ］0.8737 0.7437 0.8173 0.7549 0.9737 Polyp-Mixer［ 9 ］0.7596 0.6764 0.7954 0.7439 0.9743 Proposed 0.8849 0.7846 0.9277 0.8945 0.9835
Table 8. Experimental results of different models on the CVC-ClinicDB dataset (trained on the Kvasir-SEG dataset)
View tableTable 8. Experimental results of different models on the CVC-ClinicDB dataset (trained on the Kvasir-SEG dataset)
Method mDice mIoU Precision Recall Accuracy U-Net［ 13 ］0.6492 0.4641 0.5981 0.6594 0.9367 HarDNet-DFUS［ 25 ］0.5246 0.4397 0.6439 0.6137 0.9283 HRNetV2［ 26 ］0.7064 0.5567 0.7067 0.6819 0.9482 MSRF-Net［ 10 ］0.5943 0.4537 0.6385 0.6841 0.9519 Polyp2Seg［ 27 ］0.6492 0.5197 0.7561 0.6979 0.9424 PEFNet［ 28 ］0.6739 0.5932 0.8032 0.7138 0.9396 MSNet［ 29 ］0.5946 0.4967 0.7381 0.5738 0.9167 PraNet［ 8 ］0.7261 0.6332 0.8276 0.7194 0.9516 Polyp-Mixer［ 9 ］0.7034 0.6094 0.8409 0.7038 0.9437 Proposed 0.7271 0.6341 0.8396 0.7283 0.9512
Table 9. Experimental results of different models on the Kvasir-SEG dataset (trained on the CVC-ClinicDB dataset)
View tableTable 9. Experimental results of different models on the Kvasir-SEG dataset (trained on the CVC-ClinicDB dataset)
Method mDice mIoU Precision Recall Accuracy U-Net［ 13 ］0.5047 0.3197 0.4075 0.6537 0.7634 HarDNet-DFUS［ 25 ］0.6176 0.4739 0.8469 0.6017 0.9261 HRNetV2［ 26 ］0.5137 0.3246 0.4149 0.7614 0.7548 MSRF-Net［ 10 ］0.4873 0.3075 0.3741 0.7043 0.6832 Polyp2Seg［ 27 ］0.6061 0.5076 0.7935 0.6843 0.9274 PEFNet［ 28 ］0.5878 0.4791 0.7579 0.7184 0.8657 MSNet［ 29 ］0.5481 0.4632 0.6983 0.6542 0.8413 PraNet［ 8 ］0.5937 0.4932 0.7087 0.6172 0.8964 Polyp-Mixer［ 9 ］0.6183 0.5173 0.8142 0.7579 0.9318 Proposed 0.6259 0.5195 0.7637 0.7534 0.9183
Table 10. Results of ablation experiment
View tableTable 10. Results of ablation experiment
Method mDice mIoU GODC-U-Net 0.9293 0.8729 GODC-U-Net without GODConv 0.8628 0.8002 GODC-U-Net without SMP［ 20 ］0.9251 0.8638 GODC-U-Net without Hybrid Loss 0.9260 0.8632
Tools
Get Citation
Copy Citation Text
Gangning Lou, Peibo Sun, Shaoyao Liang, Li Zhang, Jiaqi Liu, Gangjian Hu, Liang Shen, Yongcheng Ji, Yupeng Guo. Polyp Segmentation Based on GODC-U-Net Model[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0415003
Paper Information
Category: Machine Vision
Received: Apr. 7, 2024
Accepted: Jun. 27, 2024
Published Online: Feb. 18, 2025
The Author Email:
CSTR:32186.14.LOP241038
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20