Optics and Precision Engineering, Volume. 32, Issue 4, 609(2024)
Skin lesion segmentation network with cross-attention coding
Figures & Tables(10)
Fig. 1. Schematic of overall architecture design of UH-Net network
Fig. 2. Schematic of Gated position cross self-attention mechanism
Fig. 5. Segmentation results for different networks on ISBI2017 and ISIC2018 datasets
Table 1. Results of different networks on ISBI2017 dataset
View tableView in ArticleTable 1. Results of different networks on ISBI2017 dataset
网络 IoU Dice HD95 Precesion Recall Params/M GFLOPs/G U-Net［ 6 ］75.35±0.27 83.34±0.23 41.38±0.19 93.17±0.12 75.12±0.23 31.13 55.84 UNet++［ 8 ］76.65±0.16 84.06±0.52 39.63±0.31 93.50±0.45 76.35±0.14 9.16 34.65 TransUNet［ 13 ］79.82±0.18 89.01±0.35 35.47±0.25 97.24±0.56 82.06±0.43 105.32 38.52 TransFuse［ 14 ］80.11±0.42 89.56±0.14 34.68±0.22 98.13±0.15 82.36±0.14 26.27 11.53 MedT［ 23 ］78.35±0.21 87.01±0.43 36.14±0.16 94.28±0.31 80.78±0.22 1.60 21.24 UNeXt［ 24 ］82.30±0.23 90.21±0.54 33.79±0.27 96.59±0.47 84.62±0.74 1.47 0.57 UH-Net 84.42 ±0.11 91.48 ±0.32 32.67 ±0.19 97.37 ±0.21 86.26 ±0.15 0.78 1.44
Table 2. Results of different networks on ISIC2018 dataset
View tableView in ArticleTable 2. Results of different networks on ISIC2018 dataset
网络 IoU Dice HD95 Precesion Recall Params/M GFLOPs/G U-Net［ 6 ］74.55±0.96 84.03±0.87 41.15±0.43 82.78±0.31 85.32±0.12 31.13 55.84 UNet++［ 8 ］76.12±0.65 84.96±0.71 40.08±0.51 83.56±0.13 86.41±0.26 9.16 34.65 TransUNet［ 13 ］80.51±0.72 88.91±0.63 35.21±0.47 90.01±0.36 87.84±0.23 105.32 38.52 TransFuse［ 14 ］80.63±0.32 89.27±0.13 34.42±0.33 89.61±0.11 88.93±0.18 26.27 11.53 MedT［ 23 ］79.54±0.26 87.35±0.28 35.79±0.17 87.83±0.32 86.88±0.43 1.60 21.24 UNeXt［ 24 ］81.70±1.53 89.70±0.96 34.18±0.69 88.95±0.19 90.46±0.25 1.47 0.57 UH-Net 84.12 ±0.15 91.30 ±0.24 33.24 ±0.21 92.19 ±0.42 90.52 ±0.37 0.78 1.44
Table 3. Effect of different MhGPCSA encoder settings on experimental results
View tableView in ArticleTable 3. Effect of different MhGPCSA encoder settings on experimental results
网络 分割性能/% Params/M GFLOPs/G Dice IoU Network_0 90.19±0.18 82.34±0.21 1.42 1.86 Network_1 90.72±0.32 83.12±0.45 0.78 1.42 Network_2 91.07±0.16 83.81±0.16 0.78 1.44 UH-Net 91.30 ±0.24 84.12 ±0.15 0.78 1.44
Table 4. Effect of different PosCA mechanism settings on experimental results
View tableView in ArticleTable 4. Effect of different PosCA mechanism settings on experimental results
网络 分割性能/% Params/M GFLOPs/G Dice IoU 无 90.59±0.32 82.76±0.20 0.77 1.42 +宽度注意力 90.87±0.12 82.91±0.24 0.77 1.43 +高度注意力 90.91±0.26 82.93±0.16 0.77 1.43 +全局注意力 90.83±0.18 82.86±0.29 0.77 1.43 UH-Net 91.30 ±0.24 84.12 ±0.15 0.78 1.44
Table 5. Ablation experiment
View tableView in ArticleTable 5. Ablation experiment
方法 MhGPCSA PosCA 正则化 IoU/% Dice/% HD95/% Precesion/% Recall/% Params
/M
GFLOPs/G 基线网络 - - - 80.59±0.23 88.71±0.12 35.01±0.15 82.78±0.31 86.21±0.12 1.42 1.84 √ - - 81.86±0.16 89.87±0.18 34.24±0.18 92.54±0.18 87.35±0.33 0.77 1.42 - √ - 81.38±0.35 89.31±0.98 34.81±0.27 91.76±0.52 86.99±0.26 1.42 1.86 - - √ 81.72±0.26 89.77±0.11 34.56±0.41 93.67 ±0.34 85.18±0.43 1.42 1.84 √ √ 82.46±0.12 90.35±0.21 34.01±0.26 89.94±0.20 90.76 ±0.19 0.78 1.44 √ - √ 82.76±0.20 90.59±0.32 33.93±0.18 92.38±0.16 88.87±0.21 0.77 1.42 - √ √ 82.34±0.21 90.19±0.18 34.27±0.50 92.32±0.52 88.16±0.42 1.42 1.86 UH-Net √ √ √ 84.12 ±0.15 91.30 ±0.24 33.24 ±0.24 92.19±0.42 90.52±0.37 1.44 1.44
Tools
Get Citation
Copy Citation Text
Daxiang LI, Fujie YANG, Ying LIU, Yao TANG. Skin lesion segmentation network with cross-attention coding[J]. Optics and Precision Engineering, 2024, 32(4): 609
Paper Information
Category:
Received: Aug. 24, 2023
Accepted: --
Published Online: Apr. 2, 2024
The Author Email: Fujie YANG (729378215@qq. com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20