mli0603 / ACDC2017

MICCAI ACDC 2017 Challenge

ACDC 2017

MICCAI challenge for ACDC 2017 for course project of EN.533.633 Medical Image Analysis. The proposed appoarch of using 2D AlbuNet and Random Forest Classifiers has won the competition with a test dice score 0.88 and classification accuracy of 80%. This project has also won the Best Presentation Award.

Report

See pdf

Instruction for Tensorboardx

pip install tensorboardX pip install tensorflow

start tensorboard by "tensorboard --logdir=<dir_to_store_log_file>"

Notebook

1. AlbuNet
2. UNet

Logs

1. vanilla_trained_unet_limited_data: 0.8430

2. aug_trained_unet: 0.8465 (with fine tuning)

3. UNet training performance (unfirom weight):

EPOCH 70 of 70

Training Loss: 2.4717 0 Class, True Pos 57622332.0, False Pos 230931.0, Flase Neg 135884.0, Dice score 1.0 1 Class, True Pos 481663.0, False Pos 87809.0, Flase Neg 96207.0, Dice score 0.84 2 Class, True Pos 483941.0, False Pos 156754.0, Flase Neg 167215.0, Dice score 0.75 3 Class, True Pos 541580.0, False Pos 32748.0, Flase Neg 108936.0, Dice score 0.88

4. UNet training performance (class balanced weight):

EPOCH 70 of 70

Training Loss: 3.5021 0 Class, True Pos 57608688.0, False Pos 289468.0, False Neg 149521.0, Dice score 1.00 1 Class, True Pos 460065.0, False Pos 90767.0, False Neg 117805.0, Dice score 0.82 2 Class, True Pos 465871.0, False Pos 154939.0, False Neg 185285.0, Dice score 0.73 3 Class, True Pos 533709.0, False Pos 34244.0, False Neg 116807.0, Dice score 0.88

5. UNet training performance (weight: inv(10 2 1 2)): 0.8678

EPOCH 70 of 70

Training Loss: 2.0016 0 Class, True Pos 57672216.0, False Pos 151085.0, False Neg 78722.0, Dice score 1.00 1 Class, True Pos 504550.0, False Pos 46985.0, False Neg 86626.0, Dice score 0.88 2 Class, True Pos 553362.0, False Pos 90541.0, False Neg 92723.0, Dice score 0.86 3 Class, True Pos 596317.0, False Pos 22704.0, False Neg 53244.0, Dice score 0.94

6. UNet+ResNet training performance (weight: inv(10 2 1 2) + 0.95 weight decay/epoch + 150 epochs): 0.8901

EPOCH 150 of 150

Training Loss: 0.8127 0 Class, True Pos 57693812.0, False Pos 40665.0, False Neg 37042.0, Dice score 1.00 1 Class, True Pos 566957.0, False Pos 24066.0, False Neg 27550.0, Dice score 0.96 2 Class, True Pos 623537.0, False Pos 40679.0, False Neg 34749.0, Dice score 0.94 3 Class, True Pos 635007.0, False Pos 13037.0, False Neg 19106.0, Dice score 0.98

7. UNet+ResNet training performance (prev + augmentation): 0.9286

EPOCH 150 of 150

Training Loss: 1.3366 0 Class, True Pos 57675012.0, False Pos 70108.0, False Neg 66156.0, Dice score 1.00 1 Class, True Pos 542582.0, False Pos 41528.0, False Neg 44190.0, Dice score 0.93 2 Class, True Pos 597914.0, False Pos 68496.0, False Neg 59436.0, Dice score 0.90 3 Class, True Pos 623169.0, False Pos 18956.0, False Neg 29306.0, Dice score 0.96

At epoch 132

Vaildation Loss: 2.0557 0 Class, True Pos 62978456.0, False Pos 139146.0, False Neg 90009.0, Dice score 1.00 1 Class, True Pos 532213.0, False Pos 48435.0, False Neg 98713.0, Dice score 0.88 2 Class, True Pos 595170.0, False Pos 86504.0, False Neg 72755.0, Dice score 0.88 3 Class, True Pos 605814.0, False Pos 25982.0, False Neg 38590.0, Dice score 0.95