Grading Breast Ductal Carcinomas using Spatial Activation Map Coupled Vision Transformers

Abstract

Invasive Ductal Carcinoma (IDC) is the most common form of breast cancer. Accurate grading of IDC is crucial for determining patient prognosis and treatment options. However, subtle differences between grades present a challenge for traditional classification methods. This paper proposes a novel deep-learning architecture for robust IDC grading. The proposed approach tackles the challenge of low inter-class variance by incorporating a pre-processing step followed by a feature extraction pipeline. The pipeline utilizes a sparse autoencoder to capture low-level features and a Vision Transformer (ViT) for high-level feature extraction. The extracted features are then fed into a dense neural network for classification into three distinct grades. The paper compares the performance of our proposed architecture with Convolutional Neural Networks (CNNs), transfer learning, and ViT alone. The proposed method achieves a superior accuracy of 99% on unseen test data, demonstrating its effectiveness in overcoming the limitations of traditional classification methods for grading fine-grained variations in IDC.