Web-app-Gastroscopic-Classifier

A patch-based Gastroscopic Classifier web app with Python backend using Flask micro-framework and PyTorch modified Resnet-34 Convolutional Neural Network.
This project was with Viettel Cyberspace Center (VTCC) and CNRS (Centre national de la recherche scientifique).

Application

• Assist Medical Doctors to classify gastric lesion types based on small selected patch on an endoscopy image.
• 6 Lesion types: Active Gastritis, Atrophic Gastritis, Chronic Gastritis, Intestinal Metaplasia, Normal, Ulcer.
  
  

  Sample Gastroscopic Images:

  
  

UI (User Interface)

Main screen

Lesion classification

Sample page

Dataset

• 6-class gastroscopic dataset provided by CNRS (French National Centre for Scientific Research) and 108 Military Central Hospital.
• Lesion classes: Active Gastritis, Atrophic Gastritis, Chronic Gastritis, Intestinal Metaplasia, Normal, Ulcer.

Deep Learning

• CNN model: Modified ResNet-34 model with pretrained backbones.
• Multi-perception Layer: 512-256-6.
• Regularization technique: Dropout with p=0.2.

Sample Patch

Technology used

• Backend Language: Python 3.8
• (Core Algorithm) Deep Learning framework: PyTorch
• (Core Algorithm) Image Processing library: OpenCV
• Backend Micro-framework: Flask
• Front-end: pure HTML, CSS, Javascript (no framework).