<![CDATA[Deep learning has become an important approach for medical image analysis, particularly for automated image classification. Among its various architectures, Convolutional Neural Networks (CNNs) are frequently employed to extract and distinguish visual patterns from medical images. However, the performance of a CNN may vary depending on the optimization algorithm used during the training process. This study investigates the effect of four optimization methods on different CNN architectures for the classification of Peripheral Blood Smear (PBS) images. A publicly available Kaggle dataset comprising 3,242 images and four diagnostic categories was used in the experiments. Six architectures, namely ResNet50V2, Xception, EfficientNetB2, EfficientNetB4, EfficientNetV2-S, and EfficientNetV2-M, were evaluated in combination with Adam, AdamW, NAdam, and RMSprop. Before training, the images underwent resizing, normalization, and augmentation procedures. The dataset was subsequently partitioned into training, validation, and test subsets using proportions of 80%, 10%, and 10%, respectively. Classification performance was analyzed using accuracy, loss, and confusion matrix-based evaluation. Among all tested combinations, Xception with RMSprop produced the best result, obtaining 99.69% accuracy and a loss value of 0.0194. Conversely, the lowest result was observed for ResNet50V2 with RMSprop, which achieved 93.85% accuracy. The experimental findings indicate that the optimizer can substantially influence CNN classification performance, and the effectiveness of an optimization method is dependent on the architecture with which it is combined.]]>
Copyrights © 2026
