Explainable IRViT: Inception Recurrent Vision Transformer-Based Framework for Enhanced Breast Cancer Classification with Grad CAM Analysis
Saravanan Elumalai, R Surendran
Abstract
Breast cancer is a prevalent and life-threatening disease among women worldwide that illustrates the early and accurate detection strategies to improve survival rates. Thus, this research introduces an efficient deep learning framework for automated breast cancer classification by integrating filtering based pre-processing and a novel hybrid model. Initially, the input data from a benchmark medical imaging dataset is pre-processing using a bilateral filter to eliminate noise and enhance tissue-specific features. Then, the proposed Inception Recurrent Vision Transformer (IRViT) model is utilized to classify the breast cancer, wherein the MLP layer of the vision transformer is designed with the hybrid Inception network and Recurrent model based on long short-term memory. Also, the attention mechanism of the model is enhanced with adaptive squeeze excitation based attention for choosing the most significant attributes. Thus, the proposed method is efficient in extracting spatial and sequential pattern learning from the input mammographic images. The experimental results demonstrate that the proposed model outperforms conventional models in terms of classification accuracy, precision, recall, and F1-score that illustrates its effectiveness in real-world diagnostic applications.