ENHANCING SPATIAL-TEMPORAL FEATURE FUSION WITH CONVOLUTIONAL GRUS FOR BREAST TUMOR CLASSIFICATION

Abstract

Even today, one of the main reasons for mortality among females is breast cancer, whose effective treatment needs earlier detection. In this work, a new methodology based on spatial-temporal Deep Learning for detecting and classifying breast tumors through the use of Convolutional Neural Networks (CNN) and Convolutional Gated Recurrent Unit (ConvGRU) is presented. The spatial part of the network extracts spatial features such as tumor texture, tumor shape from each slice individually, whereas the temporal part of the CNN identifies changes between different slices. The model presented in this study is capable of increasing accuracy, decreasing false positive rates, and improving robustness for tumors with irregular morphology by exploiting both spatial and temporal aspects together. For evaluation of the model, the publicly available BUSI dataset was used, and 0.9867 area under the curve (AUC), 96.88% recall, 88.57% precision, 95.40% accuracy, and 92.54% F1-score were obtained. Results have indicated the effectiveness of CNNs and ConvGRUs for automating breast cancer detection. Proposed methodology significantly surpasses baseline methods such as EfficientNetB7, DenseNet121, ConvNeXtTiny, XGBoost Ensemble, Soft Voting Ensemble in terms of classification accuracy and discriminative power. These findings support the model's applicability in clinical settings and highlight future directions in multi-class tumor classification and model optimization for deployment in resource-constrained clinical environments.