Early Prediction of Ductal Carcinoma Histological Subtypes Via Ensemble Transformer Pipelines in Digital Pathology

Breast cancer is a global concern affecting women worldwide, with invasive ductal carcinoma (IDC) being the most severe and common subtype, present in over 80 percent of cases. Traditional diagnostic methods with histological biopsies can be prone to error, with misdiagnosis rates as high as 71 percent. Deep learning models offer the potential to enhance the accuracy, reliability, and interpretability of breast cancer subtype diagnosis. Still, most existing models primarily focus on malignancy classification and lack prognostic capabilities, while also facing issues of limited generalizability and explainability that hinder their clinical implementation. To address this gap, we propose HistoFormer, an ensemble pipeline composed of high-performance vision transformers for the diagnosis of ductal carcinoma subtypes, utilizing stain-normalized histopathology images. Our approach effectively identifies relevant imaging features in histopathology scans through the use of attention-based tile mapping, enabling accurate multiclass classification. Additionally, our model showcases state-of-the-art image classification accuracies and demonstrates robust generalizability across multiple external datasets. HistoFormer presents a tool that can be potentially adapted and applied to the diagnosis and prognosis of other diseases beyond breast cancer as well, ultimately saving lives by increasing diagnostic accuracy and interpretability.