Research Article
Preliminary Study of Deep Learning Algorithms for Metaplasia Detection in Upper Gastrointestinal Endoscopy
@INPROCEEDINGS{10.1007/978-3-031-32029-3_4, author={Alexandre Neto and Sofia Ferreira and Diogo Lib\~{a}nio and M\^{a}rio Dinis-Ribeiro and Miguel Coimbra and Ant\^{o}nio Cunha}, title={Preliminary Study of Deep Learning Algorithms for Metaplasia Detection in Upper Gastrointestinal Endoscopy}, proceedings={Wireless Mobile Communication and Healthcare. 11th EAI International Conference, MobiHealth 2022, Virtual Event, November 30 -- December 2, 2022, Proceedings}, proceedings_a={MOBIHEALTH}, year={2023}, month={5}, keywords={Deep Learning Gastrointestinal Metaplasia Computer Vision Gastrointestinal Endoscopy}, doi={10.1007/978-3-031-32029-3_4} }- Alexandre Neto
Sofia Ferreira
Diogo Libânio
Mário Dinis-Ribeiro
Miguel Coimbra
António Cunha
Year: 2023
Preliminary Study of Deep Learning Algorithms for Metaplasia Detection in Upper Gastrointestinal Endoscopy
MOBIHEALTH
Springer
DOI: 10.1007/978-3-031-32029-3_4
Abstract
Precancerous conditions such as intestinal metaplasia (IM) have a key role in gastric cancer development and can be detected during endoscopy. During upper gastrointestinal endoscopy (UGIE), misdiagnosis can occur due to technical and human factors or by the nature of the lesions, leading to a wrong diagnosis which can result in no surveillance/treatment and impairing the prevention of gastric cancer. Deep learning systems show great potential in detecting precancerous gastric conditions and lesions by using endoscopic images and thus improving and aiding physicians in this task, resulting in higher detection rates and fewer operation errors. This study aims to develop deep learning algorithms capable of detecting IM in UGIE images with a focus on model explainability and interpretability. In this work, white light and narrow-band imaging UGIE images collected in the Portuguese Institute of Oncology of Porto were used to train deep learning models for IM classification. Standard models such as ResNet50, VGG16 and InceptionV3 were compared to more recent algorithms that rely on attention mechanisms, namely the Vision Transformer (ViT), trained in 818 UGIE images (409 normal and 409 IM). All the models were trained using a 5-fold cross-validation technique and for validation, an external dataset will be tested with 100 UGIE images (50 normal and 50 IM). In the end, explainability methods (Grad-CAM and attention rollout) were used for more clear and more interpretable results. The model which performed better was ResNet50 with a sensitivity of 0.75 (±0.05), an accuracy of 0.79 (±0.01), and a specificity of 0.82 (±0.04). This model obtained an AUC of 0.83 (±0.01), where the standard deviation was 0.01, which means that all iterations of the 5-fold cross-validation have a more significant agreement in classifying the samples than the other models. The ViT model showed promising performance, reaching similar results compared to the remaining models.
