Artificial intelligence applied to digestive endoscopy
Introduction: In recent years, deep learning methods have improved significantly and have been
implemented in fields such as medical imaging. Applying these techniques to digestive
endoscopy has led diagnosis rates for entities such as polyps similar or even better than humans.
Materials and methods: We trained a convolutional neural network to classify medical images into
two categories – with polyps or with normal mucosa – using about 800 images. For scalability
and accessibility reasons, the architecture was implemented into a web interface. To our
knowledge, this is the first solution to emphasize the importance of scalability and accessibility.
We developed an interface that can be used in real life scenarios and is easy to use, being web
enabled and accessible from any device. Results: Experimental results show that our solution is
feasible and can be implemented in clinical practice. The model was evaluated on the test set
and under these circumstances the final test accuracy was 100%. One limitation is the number
of images used for training. Whereas 800 images were used in total for training, only 100
contained normal mucosa and 700 contained polyps. With future research, the number of
images used will be increased and data enhancement techniques will be used, alongside with
endoscopy videos. Conclusion: In conclusion, deep learning advances can be successfully applied
to biomedical fields such as digestive endoscopy for tasks such as polyp classification, with great
potential of developing tools for medical professionals.