Fully Connected Inceptioned U-Net

Code for the UCL MSc Data Science and Machine Learning thesis.

Repository structure

• datasets/blastocyst.py contains Dataset classes that handle the data
• models/unet.py contains the baseline U-Net model
• models/incept_unet.py contains the proposed FC Inceptioned U-Net model
• utils/losses.py contains classes with custom losses
• utils/trainer.py contains class that handles model training, evaluation, and results visualisation
• train_*.py contain code to train the corresponding model and perform a grid search
• *.ipynb contain a more interactive code to train and/or evaluate the corresponding model

Data

Due to data sharing restrictions, we anonymised part of the data. However, if you wish to train and evaluate the model yourself, you can use the SFUDataset class with a publicly available dataset introduced by the Pacific Centre for Reproductive Medicine (PCRM) [1].

References

[1] Parvaneh Saeedi, Dianna Yee, Jason Au, and Jon Havelock. Automatic identification of human blastocyst components via texture. IEEE Transactions on Biomedical Engineering, 64(12):2968–2978, 2017.

Note

Due to data sharing restrictions the commit history wasn't published publicly, since it contains sensitive data.