Offloading the computational complexity of transfer learning with generic features.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Deep learning models demand substantial computational resources and time.
• Transfer learning leverages pre-trained models to mitigate computational demands without sacrificing performance.
• Conventional transfer learning relies on models trained on similar domains, incorporating numerous domain-specific features.

Purpose of the Study:

• To investigate the reduction of computational requirements in transfer learning models.
• To explore the impact of discarding domain-specific features from pre-trained models.
• To enhance breast cancer detection efficiency using optimized transfer learning.

Main Methods:

• Applied a novel transfer learning strategy to breast cancer detection.
• Utilized the Digital Database for Screening Mammography (DDSM) curated breast imaging subset.
• Evaluated performance using metrics like precision, accuracy, recall, F1-score, and computational resource usage.

Main Results:

• Discarding domain-specific features to a defined limit improved model performance.
• Achieved significant reductions in training time (approx. 12%), processor utilization (approx. 25%), and memory usage (approx. 22%).
• Demonstrated an approximate 7% increase in accuracy.

Conclusions:

• The proposed transfer learning strategy effectively reduces computational complexity.
• Optimizing feature selection in transfer learning enhances both efficiency and diagnostic accuracy.
• This method offers a promising approach for resource-constrained medical imaging applications.