Quadratic Luminance Vision Transformer Attention Network for Automated Mitosis Detection in Breast Histopathology Images

The primary limitation of CNN-based methods for mitosis detection in breast histopathological image classification is their inability to effectively extract features from potential regions of interest. This study presents a novel Quadratic Stain Luminance Normalized Vision Transformer Attention Network (QSLN-VTAN) designed to enhance feature extraction and classification, resulting in improved accuracy and precision. A feature aggregation sphere utilizing a quadratic discriminant classifier has been developed to integrate these features for the classification of images in mitosis detection. The stain normalization-based preprocessing not only enhances contrast but also preserves background luminance while ensuring robustness, significantly reducing data loss. The performance of QSLN-VTAN is assessed and compared using standard metrics, including precision, recall, accuracy, F1-score, and training time. The QSLN-VTAN demonstrated superior performance compared to other methods when evaluated on the ICPR 2012 dataset, achieving an overall mitosis detection rate of 96%, an F1 score of 93%, and a precision rate of 92%. Additionally, on the MITOSIS-ATYPIA-14 dataset, it exhibited a detection rate of 92%, an F1 score of 91%, and a precision of 96%.