Author Affiliations
1College of Information System Engineering, PLA Strategic Support Force Information Engineering University, Zhengzhou, Henan 450001, China2Department of Radiology, Henan Provincial People′s Hospital, Zhengzhou, Henan 450002, Chinashow less
Fig. 1. Structure of the RDB
Fig. 2. Dilated convolution operation of r=2 is performed in sequence. (a) First time; (b) second time; (c) third time
Fig. 3. Segmentation results of the dilated convolution. (a) Original image; (b) ground truth; (c) grid artifact
Fig. 4. Dilated convolution operations with different dilated rates in sequence. (a) r=1; (b) r=2; (c) r=3
Fig. 5. Structure of the TASD module
Fig. 6. Framework of the segmentation network
Fig. 7. Manually annotated image data
Fig. 8. Result of image enhancement. (a) Original image; (b) flip up and down; (c) flip left and right; (d) clockwise rotate 90°; (e) counterclockwise rotate 90°; (f) random zoom and rotate 1; (g) random zoom and rotate 2; (h) random room and rotate 3
Fig. 9. Statistics of the tumor size
Fig. 10. Segmentation results of our algorithm and traditional segmentation algorithm. (a) Original image; (b) region growth; (c) graph cut segmentation; (d) level set segmentation; (e) our algorithm; (g) ground truth
Fig. 11. Segmentation results of our algorithm and deep learning segmentation algorithm. (a) Original image; (b) UNet; (c) SegNet; (d) DeepLabv3; (e) FC-DenseNet; (f) our algorithm; (g) ground truth
Fig. 12. Effect of different modules on segmentation performance. (a) Original image; (b) CEL function; (c) remove RDB; (d) remove TASD module; (e) our algorithm; (f) ground truth
Algorithm | Dice | IOU | PA |
---|
Regional growth | 0.5222 | 0.3603 | 0.4623 | Graph cut | 0.5912 | 0.5034 | 0.5356 | Level set | 0.6145 | 0.5212 | 0.5629 | Ours | 0.8026 | 0.7317 | 0.7974 |
|
Table 1. Segmentation results of our algorithm and traditional algorithms
Algorithm | Dice | IOU | PA | Network parameters /M | Running speed /s |
---|
UNet[9] | 0.6562 | 0.5621 | 0.7027 | 7.76 | 0.5028 | SegNet[10] | 0.6622 | 0.5528 | 0.7102 | 1.425 | 0.5926 | DeepLabv3[11] | 0.6929 | 0.5943 | 0.7368 | 115 | 0.6584 | FC-DenseNet103[12] | 0.7736 | 0.7065 | 0.7682 | 9.26 | 0.7969 | Ours | 0.8026 | 0.7317 | 0.7974 | 2.21 | 0.5638 |
|
Table 2. Segmentation results of our algorithm and deep learning segmentation algorithm
Module | Dice | IOU | PA |
---|
CEL | FL | RDB | TASD |
---|
+ | | + | + | 0.7125 | 0.6613 | 0.7234 | | + | + | | 0.7422 | 0.6927 | 0.7549 | | + | | + | 0.7386 | 0.6822 | 0.7458 | | + | + | + | 0.8026 | 0.7317 | 0.7974 |
|
Table 3. Effect of different modules on segmentation performance