3D CT Scan Segmentation With Occupancy Network
This repo implements a 3D superresolution segmentation task for an airport baggage dataset.
Our final paper can be found here.
Model Architecture
To solve the problem of binary classification an Occupancy Network is utilize. The occupancy network is implemented with and without concatenation of the latent tensor from the encoding to the decoding path. Additionally conditional batch normalization in the decoding path can be utilized.
Original occupancy network with conditional batch normalization and without concatenation 
Occupancy network without conditional batch normalization and with concatenation
Usage
The occupancy network can be trained and/or tested, on the airport baggage dataset, by simply executing the main file.
python main.py
The following arguments can be passed to the main.py script.
| Argument | Default value | Info |
|---|---|---|
--train |
1 (True) | Flag to perform training |
--test |
1 (True) | Flag to perform testing |
--batch_size |
10 | Batch size to be utilized |
--lr |
1e-04 | Learning rate to use |
--gpus_to_use |
'0' | Indexes of the GPUs to be use |
--use_data_parallel |
0 (False) | Use multiple GPUs (num of GPUs must be a factor of the batch size) |
--epochs |
100 | Epochs to perform while training |
--use_cat |
1 (True) | One if concatenation should be utilized |
--use_cbn |
1 (True) | One if conditional BN should be utilized else normal BN is used |
--loss |
'cross_entropy' | Loss function to be utilized ('cross_entropy', 'dice' or 'focal') |
--load_model |
'None' | Path to model to be loaded |
Results
89 Dec 23, 2022
6 Aug 22, 2022
100 Dec 19, 2022
1.8k Jan 08, 2023
5 Oct 07, 2022
91 Jan 03, 2023
350 Dec 28, 2022
64 Dec 13, 2022
272 Dec 28, 2022
47 Dec 09, 2022
18 Dec 29, 2022
26 Nov 24, 2022
19 Aug 21, 2022
127 Dec 26, 2022
18 Dec 09, 2022
4 Apr 18, 2022
29 Oct 19, 2022
515 Dec 26, 2022
106 Nov 11, 2022
158 Nov 11, 2022