Hybrid Convolution-Transformer models for breast cancer classification using histopathological images

Breast cancer threatens the public health as it is
among the leading causes of women death due to unawareness
and diagnosis at the late stages. The detection of this cancer
in its early stage is decisive to decrease mortality rates . Deep
learning techniques are effective in analysis of medical images
and achieve high performance in detecting the abnormal fea-
tures, and classify them. Therefore, these methods are becoming
increasingly popular in breast cancer diagnosis. Convolutional
Neural Networks (CNNs) are commonly used for medical im-
age analysis, but Vision transformers (ViTs ) are becoming
more popular due to their excellent performance. However,
ViTs still fall behind state-of-the-art convolutional networks.
To overcome these limitations, many researchers have proposed
a new approach that combines the advantages of CNNs and
Transformers. This new approach overcomes the limitations of
each by extracting low-level features, strengthening locality, and
establishing long-range dependencies. In this study, the Hybrid
Conv-Transformer approach was used to extract features from
the BreakHis dataset of histopathological images. Coatnet and
ConvMixer models were then used to classify the images into
two binary classification based on both magnification-dependent
and magnification-independent categories. The findings indicated
that the suggested models exceeded prior models and recent deep
learning techniques on the BreakHis dataset.