A new architecture based on the Xception algorithm for pneumonia detection using medical image datasets

Abstract

This study aims to improve the reliability of pneumonia detection from chest X-ray images by addressing the instability and performance variability observed in conventional CNNs, particularly the original Xception architecture, under different training conditions. An improved Xception-based model (IXCEP) is proposed, in which the Entry, Middle, and Exit flows are redesigned using enhanced separable convolutions and skip connections. The model is evaluated on a single pneumonia dataset using both a train/validation split and 5-fold cross-validation, considering different learning rates and numbers of epochs, with additional ablation experiments to assess the contribution of each modified flow.

Experimental results show that IXCEP_full consistently outperforms the original Xception, achieving accuracy ranges of 88.1–98.1% (30 epochs, LR = 0.01), 94.5–97.9% (30 epochs, LR = 0.001), 96.0–99.8% (100 epochs, LR = 0.01), and 87.9–99.0% (100 epochs, LR = 0.001), with markedly reduced variability across folds. Ablation analysis reveals that the optimized Entry and Middle flows yield the most stable performance, reaching accuracies of 98.6–99.3%, whereas the Exit-only configuration shows higher sensitivity to training conditions. In contrast, the original Xception exhibits strong instability, with accuracy ranging from 51.4% to 93.8% across folds.

Additional results, including F1-score values of up to 99.8% and AUC values between 98.8% and 100%, supported by Friedman and Iman–Davenport statistical tests, confirm the statistical significance of the improvements. Grad-CAM visualizations further demonstrate that IXCEP focuses on clinically relevant lung regions. Overall, these findings recommend IXCEP as a more stable and reliable alternative to the original Xception for pneumonia detection from chest radiographs.