IMPACT OF IMAGE PREPROCESSING AND MODEL FINE-TUNING ON CHEST X-RAY BASED PNEUMONIA CLASSIFICATION
Despite improvements in clinical management, pneumonia is still a significant respiratory infection that demands timely radiological evaluation for early clinical intervention and minimizing preventable complications. While chest X-ray imaging is widely used as it is accessible, inexpensive, and informative for initial diagnosis, knowledge of pattern recognition is not commonplace due to subtle patterns of infection, variation between images and there is a lack of radiological skills in resource constrained areas. In this study, the impact of image preprocessing and fine-tuning on automated pneumonia categorization on chest X-ray pictures is explored. For dichotomous classification of pneumonia and normal, a transfer learning framework was designed based on the models: Xception, EfficientNetB0, and DenseNet201. Our pipeline involved image resizing, model-specific normalization, data augmentation, oversampling, deep radiographic feature extraction and fine-tuned classification layers to promote generalization. The metrics used to assess performance were based on training-validation curves, precision, accuracy, recall, F1 score and a confusion matrix. Experimental results show that Xception achieves the highest accuracy of 94.23% in the test when compared to EfficientNetB0 and DenseNet201. The results show that with proper preprocessing and selective fine-tuning, it is possible to enhance the performance of computer-aided pneumonia classification for reliability and accuracy.