Investigating Balanced Warm-Up and Gradual Unfreezing for Deep Learning-Based Mammography Classification

​

Deep learning has shown promise for breast cancer detection in mammography, but model training remains challenging because malignant cases are much less common than benign cases and pretrained models must be carefully adapted to medical images. This study investigates whether two fine-tuning strategies, balanced warm-up and gradual unfreezing, can improve mammography classification with a pretrained ResNet-101 model. Balanced warm-up first trains the model on a class-balanced subset before transferring it to the full imbalanced dataset, while gradual unfreezing progressively releases deeper network layers during training. Four settings were compared: direct fine-tuning, balanced warm-up, gradual unfreezing, and a combination of both strategies. The results showed that balanced warm-up improved initial learning on the balanced subset, but this advantage was not consistently maintained after training on the full dataset. Gradual unfreezing contributed to more stable optimization, although it did not consistently outperform direct fine-tuning in classification performance. These findings suggest that fine-tuning procedures should be evaluated as carefully as model architecture or learning rate. The study provides practical evidence for reducing blind trial-and-error in deep learning-based mammography analysis and highlights the importance of reporting detailed training strategies in medical image research.