Patchless Multi-Stage Transfer Learning for Improved Mammographic Breast Mass Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning (DL) shows promise in mammographic breast-mass classification but faces challenges with large datasets and generalization.
• Existing methods like ImageNet-based transfer learning (TL) and patch classifiers have limitations for standalone DL application.

Purpose of the Study:

• To propose a novel multi-stage TL approach for classifying mammographic breast masses as benign or malignant.
• To enhance the performance and generalizability of DL models for breast cancer diagnosis.

Main Methods:

• A multi-stage TL model was developed, pre-trained on ImageNet and cancer cell line images.
• The model was trained and validated on three public datasets (DDSM, INbreast, MIAS) and a mixed dataset.
• A patchless approach was employed, contrasting with traditional patch-based methods.

Main Results:

• Achieved an average five-fold cross-validation AUC of 1, 0.9994, 0.9993, and 0.9998 for DDSM, INbreast, MIAS, and mixed datasets, respectively.
• Demonstrated statistically significant performance improvement over patch-based methods (p=0.0029).
• Improved test accuracy by 8% (91.41% vs. 99.34%) on the INbreast dataset compared to patch- and whole image-based methods.

Conclusions:

• The proposed multi-stage TL method effectively classifies mammographic breast masses with high accuracy.
• This patchless approach addresses the need for large training datasets and reduces computational load for DL models in breast cancer screening.