DETECTING BREAST CANCER METASTASIS IN LYMPH NODE WHOLE-SLIDE IMAGES USING AN ATTENTION-DRIVEN OPTIMIZED MODEL

Abstract

Histopathology image-based breast cancer categorisation in computer-aided diagnostics is another key application where transfer learning has shown considerable potential in recent years. However, the optimal combination of different classifier heads, i.e. fully connected layers, in the form of which the features produced by pretrained convolutional neural networks become classified, remains largely unexplored despite their crucial role in improving discriminative capabilities. Therefore, the present study seeks to analyse how fully connected layer depth can affect classifier performance when using DenseNet-121, ResNet-50, and EfficientNet-B3 pretrained networks. Three levels of depth are compared in classifier heads: shallow – fully connected layer 1 (FC1), medium – fully connected layer 2 (FC2), and deep – fully connected layer 3 (FC3), thus representing increasing level of structural complexity after global average pooling. The analysis will cover several criteria for assessment including but not limited to classification boundary, effectiveness of feature transformation, and clinical utility in the form of accuracy, sensitivity, specificity, precision, F1-score, and AUC. This paper continues work in patch-based and deep learning methods for breast cancer detection. This comparative investigation into various network designs leads to some general rules for designing transfer learning architectures in the context of medical imaging, based on lessons learned from the analysis of fully connected classifier layers used for the task of predicting metastasis in lymph nodes.

Experiments reveal that deep classifier designs enable fine-grained interaction between features as well as accurate decision boundaries required for achieving clinical standards. Deep architectures can be successfully utilized together with data augmentation approaches and multiscale analysis tools. Shallow architectures suffice for the purpose of image recognition in general but cannot cope with subtle histopathological differences and class imbalance in the context of medical image analysis. The findings indicate that optimizing the classifier head emerges as an effective and computationally inexpensive strategy for enhancing performance without having to retrain backbone models or acquire new data.