DeepRad Aetos

Innovation Description
DeepRad Aetos is a prototype mobile diagnostic tool that uses an embedded deep learning model to analyse chest X-ray images (CXRs) and classify them as Normal, Pneumonia or Tuberculosis directly on a smartphone, without requiring an internet connection.
At the core of the application is a ResNet50-based convolutional neural network, trained using transfer learning on a balanced dataset of 2,100 chest radiograph images, 700 per class, drawn from publicly available Kaggle datasets. The top classification layers of the original ResNet50 were replaced with custom dense layers, dropout regularization and a softmax output optimized for three-class medical image classification. A separate Out-of-Distribution (OOD) binary classifier, also built on a frozen ResNet50 base, acts as a validation gate. It determines whether an input image is a valid chest radiograph before passing it to the diagnostic model, preventing inappropriate input from reaching the classifier.
Both models were converted to TensorFlow Lite (TFLite) format for efficient on-device inference, enabling the application to run fully offline on standard Android smartphones. The mobile application is built in Flutter and features a layered architecture: presentation, business logic, data access and model layers. It has a local SQLite database for storing and managing scan histories. Users can submit X-rays from their camera or gallery, receive a classification result with confidence scores, view and annotate their scan history and mark submissions for follow-up. The tool provides a standardized, objective classification of a chest radiograph in seconds alongside a confidence score that supports clinical decision-making. It is designed to assist and support medical personnel, not to replace expert diagnosis.
The model achieved an overall test accuracy of 96%. For tuberculosis specifically, it achieved a precision of 97% and a recall of 99% (F1 score = 0.98). For pneumonia, precision reached 98% with a recall of 93% (F1 score = 0.96).

The Problem It Solves
Respiratory diseases, particularly tuberculosis and pneumonia, are among the leading causes of death in sub-Saharan Africa and other low-income regions. Chest radiograph analysis is a cost-effective and non-invasive first-line diagnostic tool for both conditions. However, its usefulness depends entirely on access to trained radiologists or experienced clinicians who can accurately interpret the images.
In many resource-constrained settings, that access simply does not exist. Where expertise is absent or stretched thin, patients have to wait. For tuberculosis and pneumonia, that wait means missed cases, delayed treatment and continued transmission in the communities least equipped to absorb it.

Potential Impact
In Kenya and across sub-Saharan Africa, tuberculosis and pneumonia continue to claim lives that timely diagnosis could save. The bottleneck is rarely the X-ray; it's the expert who can read it. DeepRad Aetos places that capability directly in the hands of the health worker already on the ground.
Medical personnel in resource-constrained facilities can use the tool to receive an objective, confidence-scored classification result in seconds with no internet connection and no additional infrastructure required. That result supports faster triage and better-informed referrals, with the clinician retaining the final say. Earlier detection of TB reduces transmission, faster triage for pneumonia particularly in children reduces mortality. The application runs on standard smartphones already in use across the continent, adoption does not require new hardware or new habits. It meets health workers where they are.
DeepRad Aetos is proof that AI-assisted diagnosis does not have to wait for well-resourced hospitals to trickle down to the communities that need it most. It can be built for those communities from the start.