Deep Learning-Based Breast Cancer Diagnosis from Histopathology

This study investigates the detection and classification of Invasive Ductal Carcinoma (IDC) with a CNN put in to the Breast Histopathology Images (IDC) Dataset. The dataset, comprising 277,524 50x50 patches of pixel images taken from 162 full-mount slides, presents a notable class imbalance, with 198,738 IDC-negative and 78,786 IDC-positive samples. A custom CNN architecture featuring multiple convolutional and max-pooling layers, followed by dense layers and a 0.5 dropout rate, was developed to capture fine-grained spatial features. Images were used to train the model. Data generators with an early stopping callback to counteract overfitting. Analysis of training metrics revealed a consistent reduction in training loss and an increase in training accuracy. However, validation performance indicated overfitting, as validation accuracy plateaued or slightly declined while validation loss began to rise, despite continued improvements in training metrics. These findings underscore the CNN's capacity to extract suitable features for breast cancer classification while highlighting the critical need to manage class imbalance and overfitting for improved generalization. Incorporating bigger and more varied datasets, using transfer learning for generalization, and implementing the system as a real-time clinical system will be the main goals of future research.