Abstract- The performance of deep learning models in medical image classification is strongly influenced by the quality and distribution of training data. This study investigates how class imbalance and imaging noise affect model performance using chest X-ray datasets. Two transfer learning architectures, ResNet18 and MobileNetV2, were evaluated across four dataset conditions: original, imbalanced, noise-affected, and augmented. Experimental results indicate that both imbalance and noise degrade classification accuracy and sensitivity. In contrast, the use of balanced data combined with augmentation techniques significantly improves robustness, achieving an accuracy of up to 98% along with high sensitivity and specificity.
Navya K S (Thu,) studied this question.