Making Machine Learning Accessible: Efficient Transfer Learning for Mammographic Breast Cancer Classification

Breast cancer ranks as the second most impactful cancer globally, and timely detection with suitable treatment can mitigate the risk of fatalities(Muduli et al.,2022). Despite advances in mammographic disease discrimination methods, advanced multi-mammographic classification in resource-constrained healthcare environments remains challenging (Al Mansour et al.,2025), especially in regions of developing countries where GPUs are limited. This work presents a deep learning approach tailored to the CPU with an efficient class-balanced transfer learning strategy for mammographic image classification. The development and validation of this model were carried out on a set of mammographic images occupying four diagnostic classes, namely benign, bilateral, malignant, and normal images originating from a Nigerian health facility. A modified MobileNetV2 architecture with a custom preprocessing pipeline was used. The model's accuracy across all classes was 80.87%, with strong performance in malignant case detection (recall 93.94%, AUC 0.98). On standard 8GB RAM hardware, the model exhibits efficient CPU-based performance taking 47 milliseconds for preprocessing and 150 milliseconds for inference per image. Results of cross-validation analysis showed consistent performance within folds (mean validation accuracy 84.2%, SD: 1.8), which suggested robust generalization. Finally, the results suggest that sophisticated mammographic classification systems can be effectively deployed in resource-constrained clinical settings with no compromise in diagnostic accuracy.