Green Edge AI for Chest X-Ray Pneumonia Diagnosis: An EADA Benchmark

The use of clinical artificial intelligence (AI) for pneumonia detection is significantly limited by hardware constraints and the use of electricity that powers these systems in low-resource areas. This paper introduces a Green Edge AI benchmark of six CNN architectures to classify chest X-ray images using Post-Training Quantization (PTQ) and QuantizationAware Training (QAT). To explicitly quantify the trade-off between clinical efficacy and environmental impact, this paper come up with a new metric: Energy-Adjusted Diagnostic Accuracy (EADA) = F1-score/millijoule. This performance is measured on CPU only inference and calibrated on Bangladesh national grid (696 gCO2/kWh). It can be seen that the MobileNetV3-Large has the highest diagnostic quality (AUC-ROC 0.9869) and the ShuffleNetV2-x1.0 has the highest EADA (0.408), which has the
least energy consumption. QAT also lowers CO2 emissions by 27–46% over models without detectable diagnostic degradation. These lightweight models are statistically shown to have the same accuracy as their heavy counterparts, such as EfficientNet-B7, which has a carbon cost 13.5× heavier with no clinical advantage. In conclusion, INT8-quantized lightweight CNNs offer the most Pareto-efficient solution for deploying AI models in low-resource settings with a focus on climate-consciousness.