Skip to content

Latest commit

 

History

History

manuscript_data

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
img
img
 
 
README.md
README.md
 
 

README.md

Manuscript Data

1. Dataset

2. Experimental results

2.1 Evaluation on the Prostate158+ProstateX dataset for prostate anatomy segmentation

Model DSC(%) HD ASD
(baseline) 3D U-Net+DiceCELoss 81.27 19.61 2.25
3D U-Net+3DBoundaryDoULoss 81.33 19.56 2.20
3D U-Net+AWA+DiceCELoss 81.38 19.33 2.21
(ours) 3D U-Net+AWA+3DBoundaryDoULoss 81.47 19.17 2.18

2.2 Evaluation on the PI-CAI dataset for prostate lesion segmentation

Model Fold1 Fold2 Fold3 Fold4 Fold5 Mean
(baseline) ITUNet+FocalLoss 46.99 43.34 40.80 52.52 49.01 46.53
ITUNet+2DBoundaryDoULoss 45.51 44.03 45.93 55.16 54.50 49.03
ITUNet+AWA+FocalLoss 47.86 42.74 46.30 53.78 52.13 48.56
ITUNet+CZPZ+FocalLoss 49.22 46.44 46.31 53.78 50.51 49.25
(ours) ITUNet+AWA+CZPZ+2DBoundaryDoULoss 49.50 49.69 48.46 57.90 53.05 51.72

2.3 Comparison between AWA and WSO

Anatomy Segmentation :

Model DSC(%) HD ASD
3D U-Net 81.27 19.61 2.25
3D U-Net+WSO(ReLU) 80.64 19.96 2.33
3D U-Net+WSO(Sigmoid) 81.06 19.59 2.26
3D U-Net+AWA 81.38 19.33 2.21

Lesion Segmentation :

Model Fold1 Fold2 Fold3 Fold4 Fold5 Mean
ITUNet 46.99 43.34 40.80 52.52 49.01 46.53
ITUNet+WSO(ReLU) 44.59 43.13 38.12 53.01 49.09 45.59
ITUNet+WSO(Sigmoid) 45.34 42.36 40.09 54.10 49.59 46.30
ITUNet+AWA 47.86 42.74 46.30 53.78 52.13 48.56

2.4 Parameter count and FLOPs

Input Shape Model Params FlOPs
1×1×96×96×96 WSO 2 0.88M
1×1×96×96×96 AWA 0.03M 0.12G
1×1×96×96×96 3D U-Net 38.17M 15.88G
1×3×384×384 WSO 6 0.44M
1×3×384×384 AWA 0.10M 0.31G
1×3×384×384 ITUnet 18.13M 32.67G

2.5 Applicability of the AWA module in SwinUNETR and UNETR

Dataset Model DSC(%)
PI-CAI (Fold 1) (2D)SwinUNETR 48.73
PI-CAI (Fold 1) (2D)SwinUNETR+AWA 48.99
PI-CAI (Fold 1) (2D)UNETR 38.59
PI-CAI (Fold 1) (2D)UNETR+AWA 41.83
Prostate158+ProstateX (3D)SwinUNETR 80.57
Prostate158+ProstateX (3D)SwinUNETR+AWA 80.61
Prostate158+ProstateX (3D)UNETR 77.95
Prostate158+ProstateX (3D)UNETR+AWA 78.34

2.6 Applicability of the AWA module on the MSD dataset

Dataset Model Fold1 Fold2 Fold3 Fold4 Fold5 Mean
Task02_Heart (MRI) 3D U-Net 92.53 90.57 90.27 89.92 92.85 91.23
Task02_Heart (MRI) 3D U-Net+AWA 92.53 90.54 90.44 91.18 92.59 91.46
Task03_Liver (CT) 3D U-Net 76.61 72.10 72.31 72.11 75.89 73.80
Task03_Liver (CT) 3D U-Net+AWA 80.31 75.40 75.95 73.86 79.80 77.06
Task08_HepaticVessel (CT) 3D U-Net 54.53 56.32 55.26 56.73 57.43 56.05
Task08_HepaticVessel (CT) 3D U-Net+AWA 57.03 57.65 57.76 60.16 58.35 58.19
Task09_Spleen (CT) 3D U-Net 95.12 93.26 94.30 90.47 94.64 93.56
Task09_Spleen (CT) 3D U-Net+AWA 96.04 95.33 95.04 93.69 95.71 95.16

2.7 Comparison Boundary DoU Loss with other loss functions

Anatomy Segmentation :

Model Loss DSC(%) HD ASD
3D U-Net DiceCELoss 81.27 19.61 2.25
3D U-Net TverskyLoss 77.04 29.12 9.60
3D U-Net BoundaryLoss 81.47 19.17 2.18
3D U-Net BoundaryDoULoss 77.98 22.23 4.34

Lesion Segmentation :

Model Loss Fold1 Fold2 Fold3 Fold4 Fold5 Mean
ITUNet FocalLoss 46.99 43.34 40.80 52.52 49.01 46.53
ITUNet TverskyLoss 42.84 40.41 38.37 53.09 47.82 44.51
ITUNet BoundaryLoss 46.90 42.01 41.73 54.93 51.17 47.35
ITUNet BoundaryDoULoss 45.51 44.03 45.93 55.16 54.50 49.03

2.8. Hyperparameter perturbation experiments

Comparison experiments of our proposed method (ITUNet+AWA+CZPZ+2DBoundaryDoULoss) using different hyperparameters (Optimizer and Learning Rate Schedule) for prostate lesion segmentation on the PI-CAI dataset (Fold1). The evaluation metric is the Dice Similarity Coefficient (DSC).

The combination of the Adam optimizer and the CosineAnnealingLR learning rate schedule achieved better result than the result reported in the manuscript, indicating that more detailed parameter tuning can further improve the accuracy of the model.

Model Dataset Optimizer Learing Rate Schedule DSC
ITUNet+FocalLoss PI-CAI(Fold 1) Adam PolyLR 46.99
ITUNet+AWA+CZPZ+BoundaryDoULoss PI-CAI(Fold 1) Adam PolyLR 49.50
ITUNet+AWA+CZPZ+BoundaryDoULoss PI-CAI(Fold 1) SGD PolyLR 47.95
ITUNet+AWA+CZPZ+BoundaryDoULoss PI-CAI(Fold 1) Adagrad PolyLR 47.79
ITUNet+AWA+CZPZ+BoundaryDoULoss PI-CAI(Fold 1) Adam ReduceLROnPlateau 50.76
ITUNet+AWA+CZPZ+BoundaryDoULoss PI-CAI(Fold 1) Adam CosineAnnealingLR 51.03