PERFACCT

Computer Vision Based AI for Wound Detection

We developed computer vision and AI components for a wound detection system. The result was a service that segments wound areas in patient images and calculates their size based on reference markers.

Client/Company/Industry

BFI Software GmbH

Duration

26 months

Product

Service

Expertise

Software Development

Goal

The goal of the project was to develop a service for analyzing patient images. It was designed to segment wound areas and reliably calculate their size using reference markers.

Tasks

Detecting reference markers such as ChArUco boards and ColorChecker targets with OpenCV in Python
Implementing a REST web service with Flask and Gunicorn
Evaluating methods for camera calibration on mobile devices
Implementing a web service for assessing photo quality
Evaluating a neural network implementation with TensorFlow and Keras for wound segmentation
Delivering end-to-end computer vision systems for wound segmentation and OCR
Building multi-stage pipelines from detection to segmentation
Handling noise and class imbalances in real-world data
Designing and implementing versioned data pipelines with DVC
Automating experiment tracking with MLFlow
Containerizing the solution with Docker, Docker-Compose, and Kaniko
Setting up CI pipelines in GitLab CI/CD based on Docker containers
Prototyping and visualization with Jupyter Notebook
Managing the full lifecycle of real-world data through to the delivery of customer-facing APIs

Challenges

A key challenge was processing messy real-world data with noise, varying image quality, and class imbalances. In addition, multiple processing steps such as marker detection, quality assessment, and segmentation had to be integrated into robust, reproducible pipelines and provided as web services.

Programming Languages

Python

Technologies

DVC, Detectron, Docker, Docker-Compose, Flask, GitLab CI/CD, Gunicorn, Jupyter Notebook, Kaniko, Keras, MLFlow, OCR, OpenCV, REST, TensorFlow

Schematic representation of an AI-based computer vision system.

Takeaway

The project resulted in end-to-end computer vision components for wound detection and wound segmentation. Versioned data pipelines, experiment tracking, and containerized deployment established a solid foundation for further development and productive use of the solution.

Similar Projects

RIM2D - Highly Efficient 2D Hydraulic Simulation of Fluvial, Pluvial, and Urban Flooding

Hydrodynamic Simulation Web Application Geodata GPU Computing

RIM2D is an existing, highly efficient 2D hydraulic simulation model for fluvial, pluvial, and urban flooding. As part of a strategic partnership, we supported the extension of the research code with a web application and a cloud-based GPU simulation environment, enabling its transition into a market-ready product.

See project

Data Lake for Geoscience Research Data Management

Datalake Research Data Management Geodata Cloud-Native Open-Source

We developed an Open-Source S3-based data lake solution for the centralized ingestion, categorization, and searchability of data. The goal was to automate and improve manual data management through an integrated architecture with workflow orchestration, data cataloging, and access control.

See project

All Projects