ScopeAI

Overview

ScopeAI looks at colonoscopy images and flags polyps in real time, drawing a box around each one with a confidence score. The goal was a tool a small clinic could actually run, not one that needs a research lab's hardware.

The problem

Catching polyps early is how colorectal cancer gets caught early, but most detection models assume a powerful GPU. That rules out the rural clinics and small hospitals that would benefit most. I wanted something accurate that runs on limited compute.

Approach

The first prototype classified image patches with a sliding-window CNN, which was slow and clumsy. I moved to YOLOv8 Nano, which detects multiple polyps directly with bounding boxes, and trained it on the Kvasir-SEG dataset of a thousand colonoscopy images. A Flask service runs the inference; a Next.js front end handles uploads and reporting.

Key features

• Real-time analysis of single images or batches, with bounding-box detection.
• Around 6ms inference per image on modest hardware.
• Downloadable reports with annotated images and confidence scores.
• User accounts and a history of past analyses.
• 95.5 mAP50 on the held-out test set.

Tech stack

A Next.js and React front end talks to a Flask and Python backend running YOLOv8 from Ultralytics, with an early TensorFlow CNN as the baseline. Data lives in SQLite, and the two servers run side by side locally.

Outcome

ScopeAI works as a proof of concept and validated the core idea: a small, fast model can do real-time polyp detection on ordinary hardware. I wrote up how it works on the blog, and plan to open-source it.