Artificial intelligence for endoscopic grading of gastric intestinal metaplasia: another step to a greener approach on selecting individuals at risk for gastric cancer

Aims The present work was aimed at developing and evaluating the suitability of an Artificial Intelligence (AI) approach to automatically estimate the Endoscopic Grading of Gastric Intestinal Metaplasia (EGGIM) score [1].

Methods Two datasets (A and B) with a total of 533 Narrow-Band Imaging (NBI) images (153 normal, 380 with gastric intestinal metaplasia, GIM) were used. Specifically, Dataset B included full image documentation of 21 patients. All NBI images were acquired during routine clinical practice at the Portuguese Institute of Oncology in Porto from December 2019 to July 2024 and had 640x480 resolution.

Patches of 224x224 were manually selected for each image and annotated for anatomical location and EGGIM classification (0,1,2). A ResNet50 Deep Neural Network [2] was trained employing bilinear pooling through a fractal encoder. The patch-wise training was carried out using combined data from both datasets and used to learn the 3 EGGIM classes (0,1,2). Additionally, a leave-one-patient-out cross-validation was performed using dataset B. For each experience, an 80/20 data ratio was used for training and validation, respectively. The score obtained for each patch was assumed to be representative of the overall score of the corresponding anatomical location, therefore allowing to achieve a final per-patient EGGIM estimation (0-10).

Results On a per-image analysis, a balanced accuracy of 0.85±0.09, with a specificity of 0.88±0.1 and a sensitivity of 0.82±0.14 was obtained, highlighting the model’s effectiveness to classify image patches into the 3 different EGGIM classes.

The EGGIM estimation for each patient in Dataset B achieved an average error of 0.875 (out of 10) when compared to clinical experts’ evaluation and provided 100% correct clinical decisions on individual risk stratification, which are based on the previously established threshold of EGGIM≥5 for surveillance [1]. So, none of the patients requiring follow-up care would be ruled out. Surprisingly, this preliminary study suggests that patch-wise information cloud be sufficient to correctly decide on patient follow up, according to the EGGIM criteria.

Conclusions In the near future, endoscopists may use virtual chromoendoscopy and AI to easily identify GIM, enabling them to select individuals for gastric cancer (GC) surveillance with good accuracy. Following this target, we demonstrated the potential of applying AI tools in endoscopic image analyses, by automatically estimating EGGIM with high accuracy. This reinforces the utility of the EGGIM classification system in clinical practice and opens the possibility of a broader implementation of this greener (than biopsies) strategy in GC risk stratification.

Publication History

Article published online:
27 March 2025

• References

• 1 Xu Y.. et al. 'Encoding spatial distribution of convolutional features for texture representation.'. Advances in Neural Information Processing Systems 34 2021; 22732-22744
• 2 Esposito G.. et al. 'Endoscopic grading of gastric intestinal metaplasia (EGGIM): a multicenter validation study.'. Endoscopy 2019; 51 (06): 515-521