The Challenge
A brain surgeon we worked with faced scheduling challenges: patient appointments were processed strictly in the order they were booked, forcing urgent, time-sensitive cases to wait in the same queue as routine visits. While we initially built a cloud-based logistic regression scheduling tool to prioritize visits, the application failed to take hold within the hospital system and never went live in a clinical setting.
The Choice
Rather than letting the engagement end with an unused tool, we chose to push technical boundaries further on our own initiative. We set out to build an advanced, highly accurate medical imaging AI foundation that would perform reliably across diverse public datasets, run at high speeds, and execute completely on-device without relying on external cloud infrastructure or exposing patient data.
What We Built
We engineered an edge-native computer vision and MLOps framework from the ground up:
- EfficientNet-B2 Encoder Architecture: Deep neural network optimized for mobile performance and precise feature extraction from brain MRI scans.
- On-Device Mobile Inference Engine: Converted and optimized the trained model to run locally on mobile smartphones using AI on Edge techniques.
- Offline Computer Vision Pipeline: Integrated OpenCV and DICOM parsing libraries to handle medical image pre-processing directly on mobile processors.
The Outcomes
- Tech Stack: Python, EfficientNet-B2 Encoder, DICOM, Logistic Regression, MLOps Pipeline Framework, OpenCV, SVM, AI on Edge.
- Model Performance Metrics: Achieved 99.5% classification accuracy across 4 tumor classes and an 86.3% DICE score for image segmentation on held-out test datasets.
- Deployment & Privacy: Runs entirely on-device (Edge AI); no patient data or images leave the phone.
- Strategic Investment: Built as a foundational R&D asset to power future diagnostic imaging platforms—reflecting Rite Technologies' commitment to deep engineering and long-term capability building, where most of our partners rely on us for between 5 and 10 years across their evolving technical ecosystems.
The Handover
Although developed as an internal foundation rather than deployed in active clinical trials, this model established our capability to build fast, highly accurate, privacy-first computer vision tools for complex medical imaging.
Tech Stack
Python
Efficient Net B2
DICOM
Logistic Regression
MLOps
Open CV
SVM
AI on Edge
Medical Imaging
Edge AI
Health Tech
If you are exploring advanced medical imaging AI that runs securely and locally on device hardware,
let's connect.
