Home

Welcome to the jmcs-atml-bone-age-prediction wiki!

Test journal

07.05

• run transfer_learning.py (with version as of 5/6/2018) with following parameters:

1. MAIN: Chest dataset training predicts column 'Patient Age'
2. use optimizer loss 'mse'
3. Adam optimizer (for boneage training) has lr=0.0001
4. NB: only small Chest dataset was used (5.6k)

Result:
[studi5@node03 baseline]$ grep dataset transfer_learning.py.o31174.1
Chest dataset (fixed): val_mean_absolute_error: 164.20148020589627
Chest dataset (finetuning): val_mean_absolute_error: 161.06464770855962
Boneage dataset (final): val_mean_absolute_error: 33.65427841839694
NB: The training was finished at Epoch 86/250 (presumably due to convergence)

08.05

• run transfer_learning_RSNABaseline.py with chest transfer learning

Result:

• 313/313 [==============================] - 572s 2s/step - loss: 820.6397 - mae_months: 22.6683 - val_loss: 451.8546 - val_mae_months: 17.2752

Vs. baseline of Kevin:

• 313/313 [==============================] - 811s 3s/step - loss: 329.3158 - mae_months: 14.2256 - val_loss: 301.0034 - val_mae_months: 13.9117

15.05

• [Testrun#1] run transfer_learning.py with following adjustments:

1. Use full Chest dataset (112k)
2. use loss 'mae' instead of 'mse'
3. Adam optimizer (for boneage training) has lr=0.001 instead of 0.0001
4. Don't save model after every epoch (no ModelCheckpoint callback)
5. Increase patience in EarlyStopping to 10 (from 5)
6. Use entire boneage training set for training (12612 imgs.)
7. Use dedicated validation set for boneage (1426 imgs.)
8. Only 10 epochs for time reasons

Result:
Chest dataset (fixed): val_mean_absolute_error: 164.5878550175297
Chest dataset (finetuning): val_mean_absolute_error: 102.60795594894839
Boneage dataset (final): val_mean_absolute_error: 14.438980235049598

12.05

• [Testrun#1] run RSNA16BitNetServer.py

1. includes gender network

Result:

• Boneage dataset (final): val_mean_absolute_error: 39.3965670116544
• finished at Epoch 7/250

Pending Test runs:

• run transfer_learning.py with following adjustments: TODO!!!

1. MAIN: Chest dataset training uses labels 'Finding Labels' (instead of 'Patient Age')
2. Use real (official!) test set for boneage (201 imgs.)

• run transfer_learning.py with following adjustments: TODO!!!

1. Add Patient Gender Network to the solution
2. Experiment with different number of freezed layers

15.05

• Preparation of more test cases

17.05

Idea:

• Save model at end of each experiment
• lastly, load each model and evaluate against test set