Domain Adapted Few-Shot Learning for Breast Histopathological Image Classification

Abstract

The inaccessibility of high numbers of annotated data and imbalance dataset’s classes is a common phenomenon in medical microscopic image datasets. Also, collecting data from the same domain is a challenging task in microscopic imaging (e.g. same magnification factor, staining method, cancer type, imaging modality). A Conventional Deep Learning model with lack of labeled training data and domain shift could result in a misdiagnosis for such datasets. To overcome this problem, we formulated a domain adaptive few-shot learning (DA-FSL) problem and measured the perception capability of DA-FSL benchmark for medical histopathological image classification (especially for breast histopathological imaging). DA-FSL model overcomes a few-shot learning problem with domain shift by training the network on available source domain data but being able to test on target domain data with less number of available samples. DA-FSL models are validated on two publicly available breast cancer histopathological image datasets i.e., BreakHis and BreastCancer_IDC_Grade (BCIG). The presented DA-FSL framework has shown promising outcomes on both datasets. For performance evaluation we compared DA-FSL framework with other transfer learning and few-shot learning methods with different settings.