Breast Cancer Histopathological Images Classification Using Transfer Learning Combined with Separable Quaternion Moments

Breast cancer is one of the top causes of cancer deaths worldwide.Early diagnosis is an important key to improving survival rates.The execution of this task through conventional methods proves to be exceedingly timeconsuming, susceptible to errors and yields inconsistent results contingent upon the proficiency of the pathologist involved.Computer-aided diagnosis (CAD) based on Transfer Learning shows a great potential to improve diagnosis accuracy and has shown varying levels of effectiveness in histopathological images classification.This study aimed to enhance the accuracy of classifying histopathological images in both noisy and noise-free environments.In this paper, we proposed a novel approach based on transfer learning by finetuning the pretrained DenseNet201 model and combining it with separable quaternion moments.Experiments are performed using BreakHis, invasive ductal carcinoma (IDC) and BACH datasets.The BreakHis dataset demonstrates an accuracy of 97.45%, the IDC dataset shows 97.82%, and the BACH dataset achieves 98.54% of accuracy.This is attained by employing Tchebichef-Krawtchouk quaternion moments in conjunction with a finetuned DenseNet201 model, with the absence of noise.On the other hand, the proposed model shows impressive results with the presence of Gaussian, Speckle, and Salt and Pepper noise in the data.The combination of the finetuned DenseNet201 Model and separable quaternion moments enhance the accuracy of classifying histopathological images in environments with and without noise.