Transformer-Based Neural Architectures ForAutomated Cancer Classification In Histopathology Images

Abstract

Timely identification of metastatic cancer via accurate image classification is essential for enhancing patient outcomes. This research introduces a deep learning method for automated tumor identification through Transformer-Based Neural Architectures applied to histopathological images. Our model underwent training using a dataset composed of 96x96 pixel microscopic images and demonstrated remarkable performance, attaining a training accuracy of 93.9% and a validation accuracy of 93.1%. The model showed excellent effectiveness in differentiating "no tumor tissue" from "tumor tissue," reaching an ROC-AUC score of 0.9799. These findings indicate that our method is very proficient at correctly identifying tumor areas, paving the path for better diagnostic instruments in medical image analysis.