Deep Learning for the Early Detection of Invasive Ductal Carcinoma in Histopathological Images: Convolutional Neural...

Invasive ductal carcinoma (IDC) is considered the most common form of breast cancer, accounting for a significant percentage of mortality worldwide. Therefore, its early detection is vital to further improve patients' outcomes and survival rates. However, conventional diagnostic methods in the form of manual histopathological examinations are time-consuming, subjective, and prone to errors. Therefore, there is an urgent need to develop automated solutions for accurate IDC detection in histopathology images to assist pathologists in clinical decision-making.

We aim to develop and validate a convolutional neural network (CNN) model for early detection of IDC by analyzing histopathological images. The specific objectives are designing a deep learning-based technique for automated detection of IDC, assessing its performance compared to traditional diagnostic methods, and evaluating its utility in a clinical setup for early breast cancer diagnosis. These methods will be available to practitioners in underdeveloped countries via an open-source application.

The dataset for the research included 277,524 publicly available histopathological images from Kaggle, comprising both IDC-positive and IDC-negative images. About 71.6% of images were IDC-positive (class 0), while 28.4% were IDC-negative (class 1). Since our data are unbalanced, we created a weighted loss function to overcome the class imbalance problem. Further development was based on a CNN using the approach of transfer learning with a pretrained architecture called Visual Geometry Group to uplift feature extraction so that performance may improve; hence, images were preprocessed and normalized to perform augmentation with robustness. The model was developed using a split of 80% for training and 20% for testing. Model performance was measured for accuracy, sensitivity, specificity, precision, recall, and F₁-score in the confusion matrix and classification report.

From our CNN base model, we obtained an accuracy of 89% on the test set. Later, the base model was used with a weighted loss function to balance the class weights, giving a lower accuracy of 86% on the test set. Data augmentation was performed but did not improve the results. To deal with the class imbalance effectively, we performed transfer learning with a pretrained model, which gave an accuracy of 90% on the test set.

The CNN-based model thus showed accuracy and reliability for early detection of IDC from histopathological images. This technique will potentially act as an efficient and accurate assistant tool for pathologists, contributing to the early diagnosis of breast cancer and improving clinical outcomes. This paper provides an important contribution toward refining the performance of this model and widening its applications in a clinical setting by integrating it with other diagnostic techniques for better outcomes.