How to verify the performance of this model

[kingjames1155]

No description provided.

[kingjames1155]

How to test the effect of the model, such as the NCC score mentioned in the paper

[kingjames1155]

Sorry to bother you

How to visualize the final segmentation result

[fhaghighi]

How to test the effect of the model, such as the NCC score mentioned in the paper

If you want to evaluate the pre-trained Semantic Genesis on NCC (or any other target task), you first need to load the pre-trained model and then fine-tune it on the target task.
The instructions for fine-tuning Semantic Genesis on any target task can be found under the “Fine-tune Semantic Genesis on your own target task” section in Pytorch and Keras directories.

[fhaghighi]

How to visualize the final segmentation result

After you fine-tuned the pre-trained model on a segmentation target task, you can give the images as input to the target model and get the segmentation prediction; then visualize the segmentation prediction.

[kingjames1155]

How to visualize the final segmentation result

After you fine-tuned the pre-trained model on a segmentation target task, you can give the images as input to the target model and get the segmentation prediction; then visualize the segmentation prediction.

Thank you very much for your answer, and thank you for your open source code

[kingjames1155]

when i use train_autoencoder.py with luna16,Loss has been stable and does not decline，Just like the following
Have you ever encountered this problem? Have you preprocessed the dataset？

Epoch 00035: val_loss did not improve from 0.00045
Epoch 36/10000
108/108 [==============================] - 99s 912ms/step - loss: 638525.9527 - MAE: 400.1001 - MSE: 638525.8750 - val_loss: 1640486.8750 - val_MAE: 452.1590 - val_MSE: 946725.3750

Epoch 00036: val_loss did not improve from 0.00045
Epoch 37/10000
108/108 [==============================] - 97s 895ms/step - loss: 638529.2542 - MAE: 400.1557 - MSE: 638529.2500 - val_loss: 0.8887 - val_MAE: 452.6182 - val_MSE: 946749.8125

Epoch 00037: val_loss did not improve from 0.00045
Epoch 38/10000
108/108 [==============================] - 99s 916ms/step - loss: 638527.1660 - MAE: 400.1256 - MSE: 638527.2500 - val_loss: 2146421.5000 - val_MAE: 452.1830 - val_MSE: 946729.4375

Epoch 00038: val_loss did not improve from 0.00045
Epoch 39/10000
108/108 [==============================] - 98s 911ms/step - loss: 638526.7634 - MAE: 400.1210 - MSE: 638526.6875 - val_loss: 1051522.8750 - val_MAE: 452.1828 - val_MSE: 946729.1250

Epoch 00039: val_loss did not improve from 0.00045
Epoch 40/10000
108/108 [==============================] - 97s 900ms/step - loss: 638526.7060 - MAE: 400.1107 - MSE: 638526.6875 - val_loss: 0.0037 - val_MAE: 452.1874 - val_MSE: 946730.3125

Epoch 00040: val_loss did not improve from 0.00045
Epoch 41/10000
108/108 [==============================] - 98s 907ms/step - loss: 638526.3897 - MAE: 400.1137 - MSE: 638526.4375 - val_loss: 599951.1250 - val_MAE: 452.1618 - val_MSE: 946725.7500

Epoch 00041: val_loss did not improve from 0.00045
Epoch 42/10000
108/108 [==============================] - 97s 898ms/step - loss: 638526.3547 - MAE: 400.1018 - MSE: 638526.3750 - val_loss: 4.4956e-04 - val_MAE: 452.1776 - val_MSE: 946728.0000

Epoch 00042: val_loss improved from 0.00045 to 0.00045, saving model to Checkpoints/Autoencoder/Unet_autoencoder.h5
Epoch 43/10000
82/108 [=====================>........] - ETA: 22s - loss: 633764.8769 - MAE: 395.5724 - MSE: 633764.8750

[kingjames1155]

Whether I convert the data in luna16 into .npy or just still used directly,the loss does not decline with train_autoencoder.py and feature_extractor.py

[fhaghighi]

No. The loss should continuously decrease.

For preprocessing, we resample data to [1.0, 1.0, 1.0] space. We clip the data at 1000 (max) and -1000 (min) thresholds and normalize the data as (x-min(x))/(max(x) - min (x)).