Deep Learning Classification of Breast Tumours on Histopathological Images Using MobileNet-RNN and Faster R-CNN: A Comparative Experimental Study

Abstract

Background: Automated interpretation of digital histopathology may support breast-tumour classification by extracting image features that are difficult to quantify consistently by manual inspection. Deep learning architectures differ in their computational design and may therefore yield different classification performance on the same histopathology task. Objectives: To develop and compare a MobileNet-based recurrent neural network (MobileNet-RNN) and a Faster region-based convolutional neural network (Faster R-CNN) for binary classification of breast histopathology images as benign or malignant. Methods: An experimental secondary analysis was performed using the publicly available BreaKHis dataset. The source dataset comprised 7,909 histopathology images (2,480 benign and 5,429 malignant) from 82 patients. Data augmentation increased the image pool to 39,545 images (12,400 benign and 27,145 malignant). The reported split comprised 31,156 training images, 7,789 validation images, and 600 testing images. MobileNet was used for feature extraction with an RNN classifier, while Faster R-CNN used a region proposal network with convolutional feature extraction. Models were implemented in Python on Google Colab with the Adam optimizer and a learning rate of 0.01. Performance was assessed using accuracy, recall, specificity, precision, F1 score, loss, confusion matrices, and receiver operating characteristic analysis. Results: The source manuscript reported overall accuracy of 96.37% for MobileNet-RNN and 94.50% for Faster R-CNN. MobileNet-RNN achieved recall 95.23%, specificity 92.30%, precision 95.23%, F1 score 95.23%, and validation loss 0.1931. Faster R-CNN achieved recall 94.05%, specificity 94.94%, precision 95.00%, F1 score 94.52%, and validation loss 0.2102. The RNN confusion matrix contained 60 true positives, 36 true negatives, 3 false positives, and 3 false negatives (n=102), corresponding to 94.1% confusion-matrix accuracy. Faster R-CNN contained 95 true positives, 94 true negatives, 5 false positives, and 6 false negatives (n=200), corresponding to 94.5% accuracy. The reported RNN area under the ROC curve was 0.94. Conclusion: Both architectures showed high discriminatory performance for binary classification of BreaKHis histopathology images. MobileNet-RNN had the higher source-reported overall accuracy and lower validation loss, whereas Faster R-CNN showed slightly higher specificity. Because model-specific evaluation denominators differed and no external or patient-level validation was reported, the findings should be considered experimental and require validation on independent, clinically representative histopathology datasets.