Ephraim-Usani/CardioXNet-CTR-Estimation

Automated cardiothoracic ratio estimation from chest X-rays using CPU-optimized deep learning — built for Sub-Saharan Africa # CardioXNet — Automated Cardiothoracic Ratio Estimation from Chest X-rays > **CPU-optimized deep learning for cardiac screening in resource-constrained African hospitals** --- ## The Clinical Problem The Cardiothoracic Ratio (CTR) is one of the most important measurements in chest X-ray reporting. A CTR above 0.5 indicates cardiac enlargement — a critical finding in heart failure, cardiomegaly, and pericardial effusion. Traditionally, radiologists measure CTR manually by drawing lines on the X-ray image. In high-volume hospitals across Sub-Saharan Africa, this manual process: - Creates reporting delays in departments handling 100+ X-rays per day - Introduces human measurement error and inter-observer variability - Requires a trained radiologist — a scarce resource in rural Africa **CardioXNet eliminates this bottleneck entirely.** --- ## What It Does CardioXNet automatically: 1. Detects the cardiac silhouette and thoracic cavity on a chest X-ray 2. Measures the maximum cardiac diameter and maximum thoracic diameter 3. Calculates the CTR instantly and flags values above 0.5 as abnormal 4. Integrates results directly into a DICOM viewer for seamless clinical use **Result: CTR estimation in under 3 seconds. No GPU required. No radiologist needed for measurement.** --- ## Why CPU-Only Matters Most medical AI tools require expensive GPU hardware — completely impractical for district hospitals, mobile screening vans, and diagnostic centers across Nigeria and Sub-Saharan Africa. CardioXNet was built from the ground up to run on **ordinary laptops and desktop computers** — the hardware that actually exists in African hospitals. --- ## Technical Approach - **Model:** YOLOv8 Nano — chosen specifically for its CPU performance and small footprint - **Dataset:** 3,000+ chest X-rays manually annotated under radiologist supervision - **Architecture:** Custom segmentation pipeline for cardiac and thoracic boundary detection - **Integration:** DICOM viewer plu …