Optimizing lung disease classification with enhanced accuracy through transfer learning using InceptionResNetV2

For early diagnosis and efficient clinical decision-making, lung disease classification using medical imaging is essential. Nevertheless, inter-class similarity, complicated illness presentations, and the poor generalization capacity of traditional deep learning models frequently limit its effectiveness. While a number of convolutional neural networks (CNNs) have been used for the interpretation of lung images, many of them show performance instability or do not produce statistically significant gains across evaluation measures. Through a methodical comparison of cutting-edge transfer learning models, this work seeks to determine a more efficient deep learning approach for multiclass lung disease categorization. Eight popular architectures—ResNet152V2, DenseNet121, VGG16, Xception, MobileNetV3Large, EfficientNetV2L, InceptionV3, and NASNetMobile—are compared with an improved InceptionResNetV2 model. Accuracy, precision, recall, F1-score, confidence intervals, statistical significance tests, and effect size analysis are all used to evaluate performance. The suggested method obtains the greatest classification accuracy of 99.84% ± 0.15, with 98.39% precision, 98.21% recall, and an F1-score of 98.30%, according to experimental data. Strong effect sizes and considerable performance improvements over baseline models (p < 0.05) are confirmed by statistical analysis. These results show that InceptionResNetV2-based transfer learning offers a solid and dependable method for classifying lung diseases, with great promise for computer-aided clinical decision support systems.