Loading…

AboutHow It WorksModel StatsDemoTech StackGitHub →
Attention U-Net · Medical Imaging AI

AI That Sees
What Eyes
Miss.

Deep learning segmentation model trained to detect and delineate polyps in colonoscopy images — potentially saving lives through earlier detection.

76.81%
Dice Score
69.01%
IoU Score
Attention U-Net
Architecture
256 × 256
Input Size
Scroll
About This Project

Why Polyp Detection Matters

Colorectal cancer is the third most common cancer worldwide. Most cases begin as small polyps — yet up to 25% of polyps are missed during colonoscopy due to human fatigue, poor visibility, and camera angle.

This project applies Attention U-Net, a deep learning architecture that uses attention gates to focus on relevant polyp regions, trained on the CVC-ClinicDB dataset.

The goal: an AI assistant that flags regions for clinicians to review — not replacing doctors, but giving them a reliable second opinion.

~1.9M
New colorectal cancer cases per year globally
25%
Polyps missed during standard colonoscopy
90%+
Survival rate if caught at earliest stage
76.81%
Dice score achieved by this model
Pipeline

How It Works

From raw colonoscopy image to segmentation mask in under 3 seconds.

01

Upload Image

Drag and drop a colonoscopy image (JPG, PNG, or TIF). The image is sent securely to the backend API.

02

Preprocessing

The backend resizes the image to 256×256 pixels and normalises pixel values to [0, 1] for the model.

03

Attention U-Net

The model runs a forward pass. Attention gates highlight relevant polyp regions, and the decoder outputs a probability map.

04

Segmentation Mask

Pixels above 0.5 probability are classified as polyp. The binary mask is converted to an image and returned.

05

Overlay & Metrics

The polyp region is highlighted in amber on the original image. Coverage percentage and pixel count are calculated.

06

Results Displayed

Original, mask, and overlay are shown side by side. Coverage % helps gauge polyp size relative to the frame.

Performance

Model Statistics

Trained on CVC-ClinicDB with binary cross-entropy + Dice loss.

0.00%
Test Dice Score
Overlap between predicted and true mask
0.00%
Test IoU Score
Intersection over Union
~0M
Model Parameters
Trainable parameters in Attention U-Net
256×256
Input Size
RGB colonoscopy frames
Binary Mask
Output
Per-pixel polyp probability > 0.5
CVC-ClinicDB
Dataset
612 colonoscopy frames with ground truth
Live Demo

Upload. Analyse. See.

Upload any colonoscopy image and the Attention U-Net will segment the polyp region in seconds.

⚡ First request may take 60–90 s as the model loads on HuggingFace Spaces.

🔬

Results will appear here

Upload a colonoscopy image and click Analyse to see the segmentation mask and overlay.

Built With

Tech Stack

Full-stack ML deployment — from trained model to live web app, entirely free.

Next.js 14
Frontend Framework
Tailwind CSS
Styling
Three.js
3D Hero Animation
React Dropzone
File Upload UX
Flask
Backend API
TensorFlow/Keras
Model Inference
Attention U-Net
Segmentation Model
Pillow / NumPy
Image Processing
Vercel
Frontend Hosting
Render.com
Backend Hosting