Home > Projects > DeepX Health

DeepX Health

An AI-powered solution for skin cancer detection

Type

Web, Mobile, ML

Industry

Healthcare

Country

USA

Website

deepxhealth.com

Highlights

Real-time image capture and preview using the DeepX dermatoscope
A custom-trained ML model for lesion classification
Role-based access with tailored interface for dermatologists
Secure cloud storage for patient data and images

Need a high-performing solution to augment your business?

Let SoftTeco handle your tech challenges and design a software solution tailored precisely to your unique requirements.

Challenge

The client contacted SoftTeco with the need to enhance their diagnostic workflow using AI. They had an early version of a skin lesion classification model and were using a physical dermatoscopic camera device that only worked with iPads via USB-C. The key challenges included:

Developing a reliable and user-friendly iOS application to control the dermatoscopic imaging process.

Enabling image capture and transmission over a local network.

Integrating and improving upon an existing ML model for lesion classification.

Designing secure patient data storage and implementing fine-grained access control.

Solution

SoftTeco designed an web based iOS application and an AI-based classification model to assist dermatologists in identifying skin cancer at early stages. The iOS application enables clinicians to capture high-resolution dermatoscopic images, manage patient profiles, and instantly trigger lesion analysis. The AI model processes each image and returns classification results directly within the interface, helping medical professionals make more informed and accurate decisions.

Imaging
The DeepX dermatoscope, connected via USB-C to the iPad, establishes a local network with a built-in Raspberry Pi microcomputer. Two services run on this device: one manages the physical camera hardware, while the other transmits captured images over WebSocket to the iOS app.

Capture And Storage
Images are triggered via the iOS app and are displayed in real-time within a circular live preview area. Once captured, images are uploaded to the cloud and linked to specific patient profiles.

AI Evaluation
The app makes a request to the AI evaluation API, which processes the image, runs it through the classification model, and returns a risk level and lesion type. This result is shown alongside the image in the UI.

Tracking And Feedback
Clinicians can review results, compare historical data, and annotate findings. The system also logs model decisions and performance for future retraining.

Tech Stack

Backend

Node.js

Next.js

Express.js

PostgreSQL

OpenCV

Frontend

JavaScript

Angular

HTML

CSS

Redux

TypeScript

Containerization

Docker

How it Works

iOS app with embedded web interface:
We used a local web server (GCDWebServer) that runs directly within the iOS environment. This setup serves the frontend as a web app to enable seamless communication with native iOS components. At the same time, we added native iOS features: full control of the DeepX camera over USB-C and instant image transfer in real time.

Local server via GCDWebServer:
SoftTeco embedded a local HTTP server within the iOS app to allow real-time communication between the web interface and the iOS system. This enabled the app to communicate with the DeepX camera (connected via USB-C) using WebSockets.

AI-powered classification API:
We built a new computer vision model from scratch, replacing the client’s initial logic with a fully trainable ML pipeline based on EfficientNet, OTD features, and convolutional layers (CNN). The training pipeline ran on Azure Databricks with MLflow used for versioning and performance tracking.

Cloud integration and secure storage:
Patient images and metadata are uploaded to a HIPAA-compliant Azure cloud infrastructure. We implemented secure API endpoints and storage pipelines from scratch.

Role-based access and custom workflows:
The dermatologist’s user role has a specific interface and permissions, improving data segregation and patients’ safety.

Image preprocessing and feature extraction:
Prior to model inference, lesion images passed through a preprocessing pipeline that extracted numerical features from image layers, enabling more accurate classification.

Results

Delivered a fully operational AI-assisted diagnostic tool.

Integrated native iOS device hardware into a hybrid app via local networking.

Improved lesion classification performance by replacing the legacy model with a feature-rich, image-based pipeline.

Established a scalable ML infrastructure for ongoing model improvement and performance monitoring.

Our work

View all cases

SeeDoo

A computer vision device for target audience analysis

E-commerce

Blueprint

A SaaS-based business intelligence solution

Business & Finance