This repository accompanies the article entitled "Automated Classification of Oral Cancer Lesions: Vision Transformer vs Radiomics.". The project focuses on classifying cancerous oral lesions by comparing vision transformer (ViT) methods with a fully automated radiological approach. This involves the use of object detection and segmentation algorithms to effectively classifie oral lesions in cancer and control cases.
This paper is available at SSRN https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4772606 .
The logical and functional order of the project is as follows:
-
Reorganization and preprocesing (
reestructuracion.ipynb
):- This file will allow us to obtain the directories with anonymized images, separated for each set and augmented.
- It will also generate the annotation files needed for later deep learning models. Explained in more detail in the ipynb.
-
Detection (
deteccion.ipynb
):- Prediction of bounding boxes associated with the injury region using DERT, a transformer-based detection model
-
Segmentation (
segmentacion.ipynb
):- Prediction of associated masks injuries using SAM, segmentation model based on transformers.
-
Classification with ViT Models(
clasificacion_vit.ipynb
):- Use of Vision Transformer (ViT) for classification and it's analysis
- Extraction of attention maps to obtain masks.
-
Classification by Radiomics (
clasificacion_radiomica.ipynb
):- Application of radiomic techniques for classification.
This repository does not include the database used due to its character and usage restrictions. More information can be found in the associated article.
This project was developed by:
- Yolanda Vives Gilabert
- Joan Vila Francés
- José V. Bagan
- Eva Chilet Martos
We appreciate the collaboration and support of all the team members throughout the development process.
If you found this code useful, please cite our paper:
''' Chilet-Martos, Eva and Vila-Francés, Joan and Bagan, Jose V. and Vives-Gilabert, Yolanda, Automated Classification of Oral Cancer Lesions: Vision Transformers vs Radiomics. Available at SSRN: https://ssrn.com/abstract=4772606 or http://dx.doi.org/10.2139/ssrn.4772606 '''