<![CDATA[Breast tumor classification from mammogram images plays an essential role in supporting clinical decision-making, particularly because manual interpretation is often challenged by variations in breast tissue density and suboptimal image quality. This study develops a three-class classification model for normal, benign and malignant categories using the ResNet50 architecture with a transfer learning strategy on the mini-MIAS dataset, which contains 322 images with an imbalanced class distribution. Three optimizers are compared, namely Adam, RMSProp and SGD. Adam represents an adaptive moment-based optimization approach. RMSProp emphasizes stable updates under fluctuating gradients. SGD with momentum serves as a conventional baseline relying on direct gradient updates. The model is trained using a 60 percent training and 40 percent validation split with class weighting and evaluated through accuracy, AUC and F1-score metrics. Experimental results show that Adam achieves the highest performance with 68.27 percent accuracy, 88.58 percent AUC and an F1-score of 0.68. RMSProp attains 58.63 percent accuracy, 76.05 percent AUC and an F1-score of 0.59. SGD yields the lowest performance with 44.18 percent accuracy, 61.33 percent AUC and an F1-score of 0.44. Confusion matrix analysis for the Adam configuration indicates reasonably consistent recognition across all classes, although misclassification remains present. The findings demonstrate that adaptive optimizers are more effective for training ResNet50 on small and imbalanced mammogram datasets. This study provides a foundation for developing more reliable computer-aided diagnostic systems for early breast cancer detection.]]>
