Inference Engine
Upload CT scans (JPG, PNG, BMP, WEBP) to run high-speed YOLO detection tuned to an 896x896 resolution and 0.25 confidence threshold.
01 / Introduction
A full-stack diagnostic interface bridging computer vision and clinical workflows.
KidneyAI is a web-based AI assistant designed to streamline the identification of kidney stones in CT scan imagery. By replacing fragmented evaluation steps with an automated inference pipeline, it provides medical researchers and clinicians with accurate visual signals and instant report generation.
The system combines a lightweight React & TypeScript front end with a robust Flask back end, driving deep learning object detection (YOLO), persistent SQLite record archiving, and automated PDF document generation.
02 / The problem
Manual scan reviews are time-intensive and prone to reader fatigue.
Interpreting abdominal CT scans for small or subtle kidney stones requires meticulous manual review. Medical professionals often need to cross-examine multiple slices, record coordinates, log patient history, and generate diagnostic summaries across disconnected tools.
The goal was to build an integrated pipeline that accelerates object identification, standardizes confidence metrics, and instantly consolidates findings into exportable reports without friction.
03 / The solution
Automated inference, persistent storage, and clinical PDF output.
KidneyAI delivers an end-to-end workflow: CT images are validated, passed to a custom YOLO object detection network, rendered with bounding box overlays, saved to a SQLite database, and made downloadable as structured PDF reports.
Analysis History
Persist detection metrics, stone counts, bounding box coordinates, and base64-encoded annotated images inside an SQLite database.
Report Output
Generate professional A4 PDF reports via ReportLab containing scan images, confidence breakdowns, and diagnostic disclaimers.
SOLUTION / ARCHITECTURE MAP
UPLOAD → YOLO INFERENCE → SQLITE → PDF REPORT
04 / Model Architecture & Setup
Optimized computer vision fine-tuned for high precision.
Custom-Trained YOLO Network
Trained up to 100 epochs (converging at epoch 82) with an early-stopping patience of 20, optimized for CT slice contrast and feature extraction.
OpenCV Dynamic Rendering
Automated image preprocessing scales CT inputs to 896×896, draws dynamic bounding boxes with confidence scores, and outputs MIME-formatted streams.
CUDA + CPU Fallback
Seamless execution across environments with automatic hardware detection—leveraging CUDA GPU acceleration when available, with automatic CPU fallback.
05 / Application Workflow
From raw CT slice upload to saved clinical signal.
01
CT Upload
User drags & drops scan image onto the React/Vite web interface.
→
02
YOLO Inference
Flask backend processes slice through best.pt model weights.
→
03
Database Logging
Record, stone count, & confidence stored in SQLite.
→
04
PDF Report
ReportLab compiles structured summary ready for download.
SYSTEM FLOW / WORKFLOW DIAGRAM
DASHBOARD → PREDICT → SAVE → REPORT
06 / Application Modules
Clean medical-grade UI with light and dark theme support.
Designed using Tailwind CSS, TypeScript, and Lucide icons to present complex detection telemetry through intuitive dashboards, side-by-side scan reviews, and searchable history logs.
07 / System Integrity & Standards
Built for clarity, validation, and reliable performance.
Secure File Handling
Strict payload checks limit file uploads to 10 MB and validate extensions (JPG, PNG, BMP, WEBP) before OpenCV execution.
Decoupled Backend
Clean REST APIs (/predict, /history, /stats, /health) allow independent scaling of the Flask inferencing backend.
Clinical Signal Guidance
Built-in research disclaimers explicitly clarify AI detection limits for educational and demonstration purposes.
08 / Key Results & Outcomes
Accurate inference paired with seamless clinical workflow.
KidneyAI successfully demonstrates how computer vision models can be packaged into an intuitive application. The platform accelerates scan review cycles, automates record archiving in SQLite, and enables one-click PDF report generation.
Key takeaway: Combining fast deep-learning models (YOLO) with clean UI hierarchy creates an effective assistive tool for medical image evaluation.
Project details
Built by Septaneuron· Code available on GitHub.
Python
Flask
Ultralytics YOLO
OpenCV
React / TypeScript
SQLite
View the project on GitHub
Let’s build the next system.
Tell us what you are trying to make clearer, faster, or more capable.
Start a conversation