Improving Transfer Learning for Breast Cancer Detection

Background: Biopsy is widely practiced to determine if a conspicuous area in mammographic images is cancerous. Yet, 70-80% of the biopsies turn out to be noncancerous, although they may reveal benign breast diseases. Biopsy is a nontrivial procedure and may have negative side effects (fear, pain, etc...). Therefore, a reliable CAD techniques are desirable to help doctors decrease the number of unnecessary, and in retrospect negative, biopsies.

Method: The goal of this study is to develop a transfer-learning based system, which classifies benign vs. malignant tissue. We further define transfer-learning models by incorporating the following suitable preprocessing techniques: (1) phase congruency for highlighting structural and edge-type information, and (2), local texture descriptors that measure the orientional and local attributes of the textural information. We also enhance the image contrast by means of adaptive histogram equalization. These means are crucial for domain adaptation and, thereby, for improving the quality of benign/malignant classification, even when features are extracted by network trained using non-medical training set of images. We compare the performance of three methods based on their classification accuracy. Finally, feature selection via t-test is applied, to boost the classification performance.

Results: According to experimental results, the local texture description method provides the optimal preprocessing method that should be applied prior to extraction of features from VGG-19. We validate this conclusion using both breast ultrasound and mammographic images.

Conclusion: In both cases of the medical imaging datasets, selecting the efficient preprocessing technique, which describes the textural or structural information, is significant for boosting the performance of benign/malignant classification via transfer learning.