DEEP LEARNING-DRIVEN BREAST CANCER HISTOPATHOLOGICAL IMAGE CLASSIFICATION USING OPTIMIZED CNN

Abstract

Aim: To use the BreaKHis dataset to develop an ideal Convolutional Neural Networks (CNN) based system for autonomously classifying breast histopathology pictures. A number of current breast cancer classification studies have some limitations in terms of the ability to detect malignant tissues consistently, the variability of images and the over fitting that occurs. The suggested CNN approach uses convolutional layers to extract hierarchical characteristics that characterise disease from histopathology pictures in addition to image preprocessing and scaling to solve this problem. Max pooling layers were employed to minimise the dimensionality of pictures so that they could be processed effectively. Flattening or thickening can make it easier for a system to extract effective features from and classify benign versus malignant tissue. In addition, the system makes use of early stopping and dropout regularization in order to prevent over fitting; allowing the model to generalize as well. Image enhancement through augmenting images increases the variability in the image during training so that the system is more reliable and stable during the test phase. Result: In terms of classification capacity, our suggested CNN architecture greatly outperformed the baseline model for the job of classifying histopathological images. Compared to the baseline model, the accuracy of the suggested model was 1.1 percent greater. Additionally, it led to reduced loss function values, indicating improved model convergence. Additionally, the model generated an AUC-ROC value of 0.93, a higher F1-score, fewer false negatives (from 85 to 51), and a higher malignant recall (from 0.92 to 0.95). Conclusion: The proposed CNN-based system improved the accuracy of classifying breast histopathology images by enhancing the features learned in an image; by decreasing overfitting; by increasing the sensitivity to malignant cells; and by enhancing generalizability. Additionally, further study may also help provide clinicians with a better understanding of their decisions through use of XAI (Explainable AI), such as Grad-CAM.