MEE-Net: a lightweight CNN with interpretability assessment for cardiomegaly detection from chest X-rays on edge devices

Abstract Cardiomegaly is a clinically important indicator of cardiovascular risk, yet its detection from chest radiographs remains challenging due to inter-observer variability and the limited deployability of current deep learning models in resource-constrained settings. To address these limitations, we propose MEE-Net (Memory-Efficient Encoder Network), a lightweight convolutional architecture with Grad-CAM-based interpretability assessment, optimized for cardiomegaly detection on edge devices. Architecturally inspired by the encoder path of U-Net and equipped with depthwise-separable convolutions, MEE-Net contains only 83k parameters and omits batch normalization to minimize computational and memory overhead. The model was trained and evaluated on the ChestX-ray8 (5552 samples, balanced) and CheXpert (223,648 samples, imbalanced) datasets using standardized preprocessing and 5-fold cross-validation. On ChestX-ray8, MEE-Net achieved an accuracy of 0.73, $$\:F$$ 1 of 0.73, and AUC of 0.81, while on CheXpert it obtained an accuracy of 0.88, $$\:F$$ 1 of 0.86, and AUC of 0.81, showing competitive performance compared with the evaluated lightweight CNNs. Computational profiling showed the lowest CPU latency (3.5 ms per inference) and memory usage (0.5 MB) among all evaluated models. Additional evaluation on a Raspberry Pi 4 yielded an average inference latency of 23.2 ms and a runtime memory usage of 19.2 MB, further supporting the potential of MEE-Net for resource-constrained edge deployment. To quantify deployability, an integrated metric termed Edge Suitability Score ( $$\:ESS$$ ) was introduced, under which MEE-Net achieved the highest values across all six weighting schemes on ChestX-ray8 (0.89–0.94) and in five of the six schemes on CheXpert (0.70–0.84). Grad-CAM analysis suggested that MEE-Net attention maps mostly overlapped with anatomically relevant thoracic regions, with quantitative spatial alignment metrics indicating coherent correspondence between activation maps and anatomical reference masks. These findings suggest that MEE-Net provides a favorable balance between diagnostic performance and computational efficiency, while offering preliminary evidence of anatomically plausible attention behavior. Therefore, MEE-Net may represent a promising candidate for future resource-constrained cardiomegaly detection applications.