Performance Evaluation of ResNet50 and InceptionV3 Model for Breast Cancer on Histopathological Data

Abstract

One of the most commonly occurring problems in the healthcare sector is the detection of breast cancer. Timely and accurate diagnosis is essential in improving patient survival rates. This paper examines how various transfer learning backbones, such as VGG16, ResNet50, and InceptionV3 perform on the classification of histopathology of breast tissue using the BreakHis dataset. Histopathological images refer to the microscope samples that have been taken to investigate and identify the disease. The effectiveness of the individual models at distinguishing between benign and malignant tumours is strictly tested. To have reliable and generalizable findings, many preprocessing algorithms, class balancing mechanisms and data augmentation algorithms were utilized to refine overfitting, underfitting and imbalance problems among classes. Furthermore, Grad-CAM was used to improve model interpretability through the visualization of misclassification areas in the images, which were used to focus on augmentation and later fine-tuning. InceptionV3 model was the most accurate in classifications and it is rated above 97% on a test set and its success and potential can be seen in real-world diagnostic applications. Moreover, it has also been compared to state-of-the-art models.