Breast Cancer Detection using MobileNetV3: A Deep Learning Approach for Ultrasound Image Classification
Pratham Kaushik, Pooja Sharma · 2025
In this study a MobileNetV3 based deep learning model is developed for breast cancer detection on ultrasound images with the objective of reducing early diagnosis error and clinical decision making. The dataset comprises $\mathbf{1, 5 7 8}$ images categorized into three classes: Benign, normal and malignant. Data preprocessing steps (image resizing, augmentation (rotation, flipping, zooming), class weighting) were implemented in order to handle class imbalance and have efficient model training. The dataset was then split $\mathbf{8 0 \%}$ training, $\mathbf{2 0 \%}$ validation. With modifications such as extra layers to improve classification performance, MobileNetV3 was chosen due to its efficiency in handling large image datasets. Model was trained over 20 epochs with callbacks like Model-Checkpoint, ReduceLROnPlateau for learning and performance monitor. Overall the model achieves an accuracy of $\mathbf{9 2. 6 1 \%}$ with significant improvement on precision and recall for the NORMAL and Benign classes. Further areas for improvement were listed regarding the problematic misclassifications between Benign and Malignant categories. The results indicate a potential for clinical application of the model in differentiating benign from malignant cases, although better differentiation between benign and malignant cases requires optimization. Future work will consist of further improving generalization for more diverse datasets, and explore more advanced techniques, such as transfer learning and ensemble methods to improve performance in real world clinical settings.