This repository was developed for a kaggle competition UBC-OCEAN. The difficulty of this competition comes from the size of whole slide images (WSI) which are typically gigapixel.
Tellez et al., (2020) proposed neural image compression method to project WSI patches into a much smaller feature space followed by a typical CNN to classify cancer subtype given the compressed image.
Three methods of encoders are compared in my experiments.
A vanilla VAE was trained on images of 128x128 pixels.
BarlowTwins (Zbontar et al., (2021)) is selected as a framework of contrastive learning.
BarlowTwins is trained by making the cross-correlation matrix between the embeddings of two different view of images as close to the identity matrix as possible.
Network architecture and training procedure are based on BiGAN (Donahue et al., (2016)) with some modifications. Operations in the discriminator are normalized by spectral normalization (Miyato et al., (2018)). Training procedure follows TTUR (Heusel et al., (2017)) discriminator is updated twice per generator/encoder update with twice as large learning rate.
- Trained for 150 epochs with batch size
- Reconstructed images correspond to the original images
- Trained for 90 epochs with batch size 128
To handle different image size, I employed Spatial Pyramid Pooling (He et al., (2014)).
A concatenation of 1x1 max pooling (equivalent to global max pooling), 2x2 max pooling and 4x4 max pooling is used in my experiment.
Backbone of the CNN follows ConvNext (Liu et al., (2022)) architecture.
There are 5 cancer subtypes in the training data: CC, EC, HGSC, LGSC, MC. Validation set is constructed by taking one fold of StratifiedKFold splits of the training set.
Test data contains a sixth other class for rare cancer subtypes, which is not available in the training data. I did not predict other class, therefore the test balanced accuracy has 17% penalty.
Besides other class, majority (~55%) of the test set are 40x tissue microarray (TMA) whereas majority of the training set (95%) are 20x whole slide images (WSI). In addition, images in the test set are collected from different hospitals, thus quality of slides and staining techniques can be different. Those factors contribute to the discrepancy between the validation score and test score.
| encoder | validation balanced accuracy | test balanced accuracy |
|---|---|---|
| VAE | 0.50 | 0.24 |
| BarlowTwins | 0.56 | 0.18 |
| BiGAN | 0.71 | 0.21 |