Advancing Breast Cancer Diagnosis with Deep Layer-wise Feature Integration from Histopathology Images

Breast Cancer (BC) remains a leading cause of cancer-related deaths among women. Diagnosis from histopathology images can be inconsistent in pathologist assessments, highlighting the need for reliable computer-aided diagnostic systems. Existing approaches overlook multi-layer features or employ extraction methods that lack generalizability.

This study proposes a layer-wise feature integration framework to enhance BC histopathology classification.

In this analytical and computational study, using the BreakHis and ICIAR/BACH datasets, features are extracted from intermediate and high-level layers of pre-trained Convolutional Neural Networks (CNNs). Mutual Information (MI) between each layer's features and true labels is computed to identify the most informative layers, and those with the highest MI are selected. To address overfitting, Principal Component Analysis (PCA) with optimal Cumulative Explained Variance (CEV) is applied. The transformed features are then concatenated and classified using a Support Vector Machine (SVM).

The method is evaluated using VGG19, DenseNet, and ConvNextTiny across both datasets and demonstrates significant improvements in image-wise and patient-wise classification. ConvNextTiny, enhanced with Synthetic Minority Oversampling Technique (SMOTE) for class-imbalance mitigation, achieves the highest accuracy of 99.14% at 40× in BreakHis dataset.

Although the achieved accuracies of 97.32% and 99.14% (without and with imbalance handling, respectively), are slightly below the best-reported results, the method offers notable advantages in simplicity and computational efficiency. It provides a lightweight alternative to ensemble-based or fine-tuned deep models. Results highlight the critical role of intermediate layers in capturing discriminative features and challenge reliance on final-layer outputs for histopathology image classification.