Oral Squamous Cell Carcinoma Domain Shift Classification

This study was done using Python >3.11.9, please guarantee python version before running anyy scripts in order to guarantee compatibility. Research was done for the class of Computational Intelligence in Health. A summarized presenation of our results can be found in this youtube video.

Medical data takes years to collect and curate, and ideally models trained in these datasets shouldn't be used exclusively for this domain specifically. In the same laboratory, equipements can also be change, technique as well, so ideally the model needs to be able perfome well in domain shifts. Also, ideally public datasets can be used as base for problems for other labs and hospitals, not being restricted to its source. With this in mind, our goal is to see if training with weights from a previously trained model with histopathology images (transfer learning) increases the perfomance of the model. We will be using as base the setup found in Beatriz Matias Santana Maia, Maria Clara Falcão Ribeiro de Assis, Leandro Muniz de Lima, Matheus Becali Rocha, Humberto Giuri Calente, Maria Luiza Armini Correa, Danielle Resende Camisasca, Renato Antonio Krohling, Transformers, convolutional neural networks, and few-shot learning for classification of histopathological images of oral cancer, Expert Systems with Applications, Volume 241, 2024, 122418, ISSN 0957-4174, https://doi.org/10.1016/j.eswa.2023.122418.

📥 Downloading data

The dataset used for training was NDB-UFES: An oral cancer and leukoplakia dataset composed of histopathological images and patient data, curated by Maria Clara Falcão Ribeiro de Assis, Leandro Muniz de Lima, Liliana Aparecida Pimenta de Barros, Tânia Regina Velloso, Renato Krohling, Danielle Camisasca. Published in March 16, 2023. Mendeley Data, V4 , doi: https://doi.org/10.17632/bbmmm4wgr8.4.

In order to use the images, the directory data was created at the roo of the repositoruy. The folder images and the file ndb-ufes.csv obtaind from mendeley was moved to data. For the patches images, the data used was obtained by Labcin instead of the format available at mendeley, this is so that the original image names can be used. The original image names contain the parent image name, having this information is vital to guarantee that no contamination is made during train/test splits.

The data directory contained the following format:

data
|---- ndb-ufes
|     |---- images
|     |      |-- image.png
|     |      |-- image.png
|     |      |-- ...
|     |---- patches
|     |      |-- carcinoma
|     |      |    |-- image.png
|     |      |    |-- image.png
|     |      |    |-- ....png
|     |      |-- no_dysplasia
|     |      |    |-- image.png
|     |      |    |-- image.png
|     |      |    |-- ....png
|     |      |-- dysplasia
|     |      |    |-- image.png
|     |      |    |-- image.png
|     |      |    |-- ....png
|     |---- ndb-ufes.csv
|---- rahman
|     |---- First Set
|     |      |-- 100x Normal Oral Cavity Histopathological Images
|     |      |    |-- image.jpg
|     |      |    |-- image.jpg
|     |      |    |-- ....jpg
|     |      |-- 100x OSCC Histopathological Images
|     |      |    |-- image.jpg
|     |      |    |-- image.jpg
|     |      |    |-- ....jpg
|     |---- Second Set
|     |      |-- 400x Normal Oral Cavity Histopathological Images
|     |      |    |-- image.jpg
|     |      |    |-- image.jpg
|     |      |    |-- ....jpg
|     |      |-- 400x OSCC Histopathological Images
|     |      |    |-- image.jpg
|     |      |    |-- image.jpg
|     |      |    |-- ....jpg

The dataset used for test was Histopathological imaging database for Oral Cancer analysis, curated by Tabassum Yesmin Rahman, Lipi B. Mahanta, Anup K. Das, Jagannath D. Sarma. Published in January 9, 2023. Mendeley Data, V1, doi: 10.17632/ftmp4cvtmb.2

🔨 Configuring environment

Before installing dependencies, create and start your virtual environment.

python -m venv {virtual_env_name}

on linux run the command:

source {virtual_env_name}/bin/activate

on window run the command:

{virtual_env_name}\Scripts\activate

Before installing dependencies via terminal, guarantee that the tag {virtual_env_name} appears before directory path. Once confirmed, run the command:

pip install -r requirements.txt

PyTorch was used in this project. While it's included in the requirements, I recommend following PyTorch's how to install guide, substituing with the versions included in the requirements. This way you can guarantee it's compatible with your CUDA version. For example:

pip install torch==2.3.1 torchvision==0.18.1 --index-url https://download.pytorch.org/whl/cu118

Chapters

Research phase

1. Research about domain shift
2. Research about SOTA image classifiers
3. Research about domain shift applied to histopathology images
4. Research about OSCC image classification

Experimental phase

Base model: DenseNet121

Part 1

1. Split datasets in test and train for both Rahman and P-NDB-UFES.
2. Run training experiments with patches dataset with the same transformation
3. Run training experiments with Rahman dataset with the same transformation

Part 2

4. Test P-NDB-UFES model with P-NDB-UFES test set
5. Test P-NDB-UFES model with Rahman test set (no training done on Rahman here)
6. Test Rahman model with Rahman test set
7. Test Rahman model with P-NDB-UFES test set (no training done on P-NDB-UFES here)

Part 3

8. Training P-NDB-UFES model on Rahman training set (transfer learning) 100x + 400x
9. Test P-NDB-UFES-Rahman model with Rahman test

Part 4

10. Analyze results

Run experiments

sh part1_train.sh

Then you have to change part2_train.sh variables to your best model.

sh part2_test.sh

Then you have to change part3_train.sh variables to your best model.

sh part3_train.sh

Then you have to change part3_test.sh variables to your best model.

sh part3_test.sh

Authors

• Beatriz Matias Santana Maia
• Vitor Bonella
• Higor David Oliveira

References

1. Stacke K, Eilertsen G, Unger J, Lundstrom C. Measuring Domain Shift for Deep Learning in Histopathology. IEEE J Biomed Health Inform. 2021 Feb;25(2):325-336. doi: 10.1109/JBHI.2020.3032060. Epub 2021 Feb 5. PMID: 33085623.