Multiscale Residual Attention Feature Fusion Network for Ovarian Cancer Subtype Classification

Abstract

The classification of ovarian cancer into specific subtypes is essential for distinguishing tumor categories based on cellular morphology and tissue architecture observed in histopathological images. However, subtype classification remains challenging due to morphological similarities, variations in tissue patterns and limited discriminative feature representations. To solve these challenges, this research proposes a Multiscale Residual Attention Feature Fusion Network (MRAFF-Net) is introduced for automatic ovarian cancer subtype classification. The algorithm integrates multiscale feature extraction, residual learning, channel–spatial attention and adaptive feature fusion. First, histopathological images are resized and normalized, followed by parallel feature extraction at fine, intermediate and coarse scales. Residual connections preserve informative representations, while channel and spatial attention mechanisms emphasize discriminative tissue features. Subsequently, adaptive feature fusion combines complementary multiscale representations into a unified feature vector, which is passed to fully connected layers and a Softmax classifier to predict the ovarian cancer subtype. Performance is evaluated using accuracy, precision, recall, F1-score, specificity and the area under the Receiver Operating Characteristic Curve (ROC-AUC). The experimented results indicate an accuracy of 96.82% under the evaluated experimental conditions. Overall, MRAFF-Net integrates multiscale histopathological feature extraction and feature fusion for ovarian cancer subtype classification. Further evaluation using larger, independent, multicenter cohorts is required to assess generalizability and clinical applicability.