https://github.com/drawcodeboy/scoliosis_seg-task

https://github.com/drawcodeboy/scoliosis_seg-task

Last synced: about 1 year ago
JSON representation

• Host: GitHub
• URL: https://github.com/drawcodeboy/scoliosis_seg-task
• Owner: drawcodeboy
• Created: 2024-07-26T05:16:04.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2024-12-20T04:12:56.000Z (about 1 year ago)
• Last Synced: 2025-01-05T23:14:15.371Z (about 1 year ago)
• Language: Python
• Size: 327 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
* SegFormer Codebase

* SegNeXt Codebase

## When Distributed Learning (2 GPUS)

### SegFormer

* 0, 1, 2 = batch-size 16

* 3, 4 = batch-size 8

* 5 = batch-size 4

### SegNeXt

* T, S, B = batch-size 16

* L = batch-size 8

## Scoliosis Results

* SegFormer-B0

    * Params: 3.7M

    * GFLOPS: 13.1552G

    * It was trained 72 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.8782 | Dice: 0.9346 | Precision: 0.9360 | Recall: 0.9357 | loss: 0.067758

    * Test Time: 00m 02s

* SegFormer-B1

    * Params: 13.7M

    * GFLOPS: 25.7578G

    * It was trained 87 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.8916 | Dice: 0.9422 | Precision: 0.9413 | Recall: 0.9453 | loss: 0.059165

    * Test Time: 00m 03s

* SegFormer-B2

    * Params: 27.3M

    * GFLOPS: 98.3070G

    * It was trained 51 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.8960 | Dice: 0.9447 | Precision: 0.9406 | Recall: 0.9509 | loss: 0.056033

    * Test Time: 00m 05s

* SegFormer-B3

    * Params: 47.2M

    * GFLOPS: 126.2161G

    * It was trained 79 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.9072 | Dice: 0.9509 | Precision: 0.9486 | Recall: 0.9552 | loss: 0.049237

    * Test Time: 00m 06s

* SegFormer-B4

    * Params: 64.0M

    * GFLOPS: 154.6397G

    * It was trained 82 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.9087 | Dice: 0.9517 | Precision: 0.9516 | Recall: 0.9538 | loss: 0.048405

    * Test Time: 00m 07s

* SegFormer-B5

    * Params: 84.6M

    * GFLOPS: 181.1287G

    * It was trained 58 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.9078 | Dice: 0.9512 | Precision: 0.9495 | Recall: 0.9548 | loss: 0.048933

    * Test Time: 00m 08s

* SegNeXt-T

    * Params: 4.8M

    * GFLOPS: 16.5449G

    * It was trained 24 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.8679 | Dice: 0.9288 | Precision: 0.9229 | Recall: 0.9372 | loss: 0.071246

    * Test Time: 00m 02s

* SegNeXt-S

    * Params: 14.6M

    * GFLOPS: 31.7697G

    * It was trained 12 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.8710 | Dice: 0.9306 | Precision: 0.9285 | Recall: 0.9351 | loss: 0.069426

    * Test Time: 00m 03s

* SegNeXt-B

    * Params: 27.6M

    * GFLOPS: 53.2278G

    * It was trained 96 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.8830 | Dice: 0.9374 | Precision: 0.9364 | Recall: 0.9404 | loss: 0.062625

    * Test Time: 00m 05s

* SegNeXt-L

    * Params: 46.2M

    * GFLOPS: 87.3279G

    * It was trained 93 EPOCHS

    * Evaluate: 100.00%

    * IoU: 0.9005 | Dice: 0.9472 | Precision: 0.9460 | Recall: 0.9505 | loss: 0.052768

    * Test Time: 00m 07s

## ICH_only Results

### threshold = 0.5

* SegFormer-B0

    * Params: 3.7M

    * GFLOPS: 13.1552G

    * It was trained 35 EPOCHS

    * IoU: 0.0848 | Dice: 0.1427 | Precision: 0.0977 | Recall: 0.4498 | loss: 0.887513

* SegNeXt-T

    * Params: 4.8M

    * GFLOPS: 16.5449G

    * It was trained 53 EPOCHS

    * IoU: 0.3833 | Dice: 0.4889 | Precision: 0.6041 | Recall: 0.4822 | loss: 0.519992

* SegNeXt-S

    * Params: 14.6M

    * GFLOPS: 31.7697G

    * It was trained 30 EPOCHS

    * IoU: 0.3950 | Dice: 0.5020 | Precision: 0.7385 | Recall: 0.4474 | loss: 0.551906

* SegNeXt-B

    * Params: 27.6M

    * GFLOPS: 53.2278G

    * It was trained 22 EPOCHS

    * IoU: 0.3648 | Dice: 0.4579 | Precision: 0.7270 | Recall: 0.4020 | loss: 0.591055

* UNet-Base

    * Params: 31.0M

    * GFLOPS: 341.0395G

    * It was trained 51 EPOCHS

    * IoU: 0.3787 | Dice: 0.4816 | Precision: 0.4092 | Recall: 0.8859 | loss: 0.804174

* SegFormer-B0 (Data Augmentation)

    * It was trained 467 EPOCHS

    * IoU: 0.0250 | Dice: 0.0473 | Precision: 0.0250 | Recall: 0.9317 | loss: 0.957689

### threshold = 0.6

* SegFormer-B0

    * It was trained 35 EPOCHS

    * IoU: 0.0850 | Dice: 0.1429 | Precision: 0.0994 | Recall: 0.4282 | loss: 0.887513

* SegNeXt-T

    * It was trained 53 EPOCHS

    * IoU: 0.3830 | Dice: 0.4883 | Precision: 0.6068 | Recall: 0.4798 | loss: 0.519992

* SegNeXt-S

    * It was trained 30 EPOCHS

    * IoU: 0.3945 | Dice: 0.5014 | Precision: 0.7435 | Recall: 0.4465 | loss: 0.551906

* SegNeXt-B

    * It was trained 22 EPOCHS

    * IoU: 0.3643 | Dice: 0.4573 | Precision: 0.7277 | Recall: 0.4011 | loss: 0.591055

* UNet-Base

    * It was trained 51 EPOCHS

    * IoU: 0.3916 | Dice: 0.4975 | Precision: 0.4250 | Recall: 0.8769 | loss: 0.804174

### threshold = 0.7

* SegFormer-B0

    * It was trained 35 EPOCHS

    * IoU: 0.0855 | Dice: 0.1433 | Precision: 0.1015 | Recall: 0.4073 | loss: 0.887513

* SegNeXt-T

    * It was trained 53 EPOCHS

    * IoU: 0.3828 | Dice: 0.4878 | Precision: 0.6097 | Recall: 0.4773 | loss: 0.519992

* SegNeXt-S

    * It was trained 30 EPOCHS

    * IoU: 0.3944 | Dice: 0.5011 | Precision: 0.7450 | Recall: 0.4456 | loss: 0.551906

* SegNeXt-B

    * It was trained 22 EPOCHS

    * IoU: 0.3637 | Dice: 0.4566 | Precision: 0.7279 | Recall: 0.4002 | loss: 0.591055

* UNet-Base

    * It was trained 51 EPOCHS

    * IoU: 0.4009 | Dice: 0.5085 | Precision: 0.4374 | Recall: 0.8683 | loss: 0.804174

### threshold = 0.8

* SegFormer-B0

    * It was trained 35 EPOCHS

    * IoU: 0.0858 | Dice: 0.1435 | Precision: 0.1040 | Recall: 0.3829 | loss: 0.887513

* SegNeXt-T

    * It was trained 53 EPOCHS

    * IoU: 0.3821 | Dice: 0.4867 | Precision: 0.6134 | Recall: 0.4742 | loss: 0.519992

* SegNeXt-S

    * It was trained 30 EPOCHS

    * IoU: 0.3934 | Dice: 0.5001 | Precision: 0.7455 | Recall: 0.4441 | loss: 0.551906

* SegNeXt-B

    * It was trained 22 EPOCHS

    * IoU: 0.3632 | Dice: 0.4559 | Precision: 0.7287 | Recall: 0.3991 | loss: 0.591055

Test Time: 00m 03s

* UNet-Base

    * It was trained 51 EPOCHS

    * IoU: 0.4098 | Dice: 0.5188 | Precision: 0.4503 | Recall: 0.8577 | loss: 0.804174

### threshold = 0.9

* SegFormer-B0

    * It was trained 35 EPOCHS

    * IoU: 0.0865 | Dice: 0.1440 | Precision: 0.1085 | Recall: 0.3500 | loss: 0.887513

* SegNeXt-T

    * It was trained 53 EPOCHS

    * IoU: 0.3815 | Dice: 0.4856 | Precision: 0.6180 | Recall: 0.4696 | loss: 0.519992

* SegNeXt-S

    * It was trained 30 EPOCHS

    * IoU: 0.3926 | Dice: 0.4991 | Precision: 0.7469 | Recall: 0.4424 | loss: 0.551906

* SegNeXt-B

    * It was trained 22 EPOCHS

    * IoU: 0.3620 | Dice: 0.4544 | Precision: 0.7290 | Recall: 0.3971 | loss: 0.591055

Test Time: 00m 03s

* UNet-Base

    * It was trained 51 EPOCHS

    * IoU: 0.4210 | Dice: 0.5315 | Precision: 0.4678 | Recall: 0.8406 | loss: 0.804174

## ICH_all Results

* 해당 방법론은 어려운 이유: 학습이 끝나면 SegFormer는 괜찮으나 SegNeXt는 모두 Negative라 판단하는 경향이 있다.

* SegFormer-B0

    * It was trained 51 EPOCHS

    * IoU: 0.3809 | Dice: 0.3911 | Precision: 0.4172 | Recall: 0.9365 | loss: 0.951504

* SegNeXt-T (??)

    * It was trained 1 EPOCHS

    * IoU: 0.8880 | Dice: 0.8880 | Precision: 1.0000 | Recall: 0.8880 | loss: 0.111111

Test Time: 00m 09s