Author Affiliations
1School of Automation, Nanjing University of Information Science & Technology, Nanjing 210044, Jiangsu , China2Jiangsu Atmospheric Environment and Equipment Technology Collaborative Innovation Center, Nanjing 210044, Jiangsu , Chinashow less
Fig. 1. Overall structure of the improved model
Fig. 2. RR block. (a) Overall structure, the residual block is nested with the recombination block, and the recombination block is nested within the SegSE block; (b) structure of the recombination block; (c) structure of the recalibration block; (d) structure of the residual block
Fig. 3. Module structure. (a) SE module; (b) recalibration module
Fig. 4. Attention mechanism structure
Fig. 5. Flow chart of pretreatment
Fig. 6. Broken line diagram of the influence of RR block on the model. (a) Dice Score; (b) Sensitivity
Fig. 7. Segmentation results. (a) CT original image; (b) gold standard; (c) prediction result; (d) comparison between gold standard and prediction result
Fig. 8. Three-dimensional diagram of segmentation results, the left side of each image is the prediction result, and the right side is the comparison between the segmentation result and the real annotation. (a) MPUNet; (b) nnU-Net; (c) Spider UNet; (d) proposed method
Module | Channel | Dice /% | Sen /% | Spe /% | Acc /% |
---|
+RR block | [48,64] | 60.83 | 67.97 | 97.92 | 98.34 | [48,64,96] | 63.95 | 69.72 | 98.93 | 99.12 | [48,64,96,144] | 61.43 | 71.32 | 98.85 | 98.91 | [48,64,96,144,192] | 55.97 | 67.92 | 97.73 | 97.87 | -RR block | [48,96] | 30.13 | 33.11 | 93.24 | 93.63 | [48,96,128] | 37.21 | 48.65 | 94.43 | 95.12 | [48,96,128,256] | 41.22 | 51.23 | 95.53 | 96.14 | [48,96,128,256,512] | 33.85 | 41.91 | 94.21 | 94.95 |
|
Table 1. Quantitative analysis of the influence of the RR block on the model
Loss | Attention | RR block | Dice /% | Sen /% | Spe /% | Acc /% |
---|
Tversky Loss | | | 41.22 | 51.23 | 95.53 | 96.14 | + | | 42.21 | 52.33 | 96.11 | 97.34 | | + | 63.95 | 69.72 | 98.93 | 99.12 | + | + | 64.80 | 70.51 | 99.12 | 99.23 | GD Loss | | | 40.22 | 52.31 | 95.22 | 96.81 | + | | 41.95 | 53.63 | 96.61 | 97.53 | | + | 62.53 | 71.81 | 98.60 | 98.82 | + | + | 63.72 | 73.15 | 99.31 | 99.10 |
|
Table 2. Influence of each block and loss function on the model
Model | Year | Dice Score |
---|
MPUNet[19] | 2019 | 59.00 | UMCT[20] | 2020 | 63.00 | nnU-Net[21] | 2019 | 63.00 | Spider UNet[22] | 2021 | 45.60 | C2FNAS-Panc[23] | 2020 | 64.30 | Proposed model | | 64.80 |
|
Table 3. Segmentation effect of different models