CoaT-RADS: Leveraging C-Scale Conv-attentional Transformers For Automated Mammography Classification

Abstract

Breast cancer is one of the leading causes of mortality among women worldwide, and early diagnosis plays a critical role in improving survival rates. Globally, approximately 2.3 million new cases are reported each year, with nearly 685,000 deaths largely attributed to late-stage detection. In India, breast cancer constitutes more than 27% of all female cancers, emphasizing the urgent need for efficient and reliable diagnostic solutions. With the expansion of screening programs, automated mammography analysis has become increasingly important in hospitals, diagnostic centers, teleradiology services, and rural healthcare settings where access to expert radiologists is limited. Conventional manual interpretation of mammograms is often affected by inter-observer variability, fatigue-related errors, and challenges in identifying subtle abnormalities, especially in dense breast tissues with overlapping structures. This research presents a structured mammography evaluation framework that integrates Co-Scale Conv-Attentional Image Transformers (CoaT) with feature-driven representations derived from Extreme Gradient Boosting (XGB), Histogram Gradient Boosting (HGB), and Fast Interpretable Oblique Trees (FIOT). Each representation contributes uniquely to understanding mammographic patterns without depending solely on automated inference. XGB organizes features using gradient-based tabular structures, capturing variations in tissue density and lesion boundaries, while HGB represents grayscale and texture information through histogram-based distributions for improved intensity analysis. FIOT employs an oblique-tree structure using linear feature combinations, enabling enhanced interpretability of tissue irregularities. Comparative analysis within the COAT framework indicates that FIOT achieves superior performance due to its stable and interpretable feature interactions, followed by HGB, whereas XGB demonstrates comparatively lower consistency.