Chest X-rays are critical to diagnosing lung diseases across Africa, yet hospitals often face enormous pressure: too many patients, too few radiologists, and limited capacity for timely diagnoses. To ease this burden, we’ve built a simple but effective two-stage AI system that fits the realities of overstretched clinics.
The first step filters out normal scans using EfficientNet-B0, a compact model we found to be the best compromise between speed and accuracy after testing it against MobileNetV3-Small and ResNet-18. To prevent the model from being skewed by class imbalance, we combined weighted loss functions with balanced sampling.
If a scan looks suspicious, it’s passed to a second model trained to recognize 14 chest conditions. For this, we tested DenseNet-121, ResNet-50, and EfficientNet-B2. All training was done using the NIH ChestX-ray dataset. To track learning progress, we plotted AUROC scores as training advanced.
The final system outputs structured reports and uses Grad-CAM to show which image regions influenced predictions. This helps clinicians understand and trust the results. Everything in the pipeline uses open-source tools and public datasets—no black boxes, no paywalls. It’s designed to work where it’s needed most: busy hospitals, TB clinics, and maternal care wards across the continent.
Keywords: chest X-ray; deep learning; medical imaging; clinical AI; model interpretability; open source; Grad-CAM; AUROC; NIH dataset; diagnostic workflow; maternal health; TB screening
