Deep Learning and Generative AI in Helicobacter Pylori Diagnosis: Image Enhancement, Classification and Medical Chatbot

Yauri L.; Romero J.; Tovar N.; Mondalgo J.; Camarena L. · C3 2025 - IEEE Colombian Caribbean Conference (2025) · DOI: 10.1109/c366505.2025.11340323

Resumen

Helicobacter pylori is a bacterium associated with severe gastric diseases such as ulcers and gastric cancer. Early detection is critical, but conventional methods are often slow and dependent on specialists. In this study, an artificial intelligence and generative approach was proposed to assist in the diagnosis of Helicobacter pylori from digitized gastric biopsies. The research framework included: (1) Image resolution enhancement using Real-ESRGAN, (2) structured segmentation of images into 512 × 512 px patches, (3) training of classifiers based on DenseNet121 and InceptionV3 using transfer learning techniques, and (4) evaluation through standard binary classification metrics. Additionally, synthetic image generation was explored using StyleGAN3, and Ronsoco Bot was developed a medical chatbot powered by large language models (LLMs) to provide evidence based automated clinical support. Six models were trained on a dataset of 8,998 images (90 %) and evaluated on a test set of 1,002 images (10%), which included both positive and negative Helicobacter pylori cases. The best-performing model, named Ronsoco, achieved an accuracy of 0.859, a precision of 0.868, a recall of 0.862, a F1-score of 0.860 and an AUC-ROC of 0.934. Regarding the conversational component, Ronsoco Bot achieved a context recall of 0.833, answer relevancy of 0.794 and faithfulness of 0.907. Limitations were identified due to computational resource constraints and data scarcity. In future work, we propose optimizing generative models, integrating clinical metadata and validating the solution in hospitals, with potential to extend this approach to other gastrointestinal pathologies using generative AI and deep classification.

Instituciones: Universidad Nacional Mayor de San Marcos