MLFRNet: Multi-Scale Local Feature Refinement Network Using Multi-Scale Depthwise Convolution for HER2 Score Classification

Background/Objectives: Human Epidermal Growth Factor Receptor 2 (HER2) is a key biomarker for selecting treatment strategies in breast cancer patients. In routine pathology workflows, HER2 assessment is performed by pathologists through examination of immunohistochemistry (IHC)-stained sections under a microscope. Manual IHC assessment requires expert interpretation and can be subject to inter-observer variability, which motivates the study of computer-assisted analysis in digital pathology. Methods: MLFRNet uses a ConvNeXt-Tiny backbone with a Multi-scale Local Feature Refinement (MLFR) module that integrates Efficient Channel Attention (ECA), a multi-scale depthwise convolution block (MSB), and a residual refinement block (RB). ECA reweights channel representations without dimensionality reduction, after which MSB applies parallel depthwise convolutions with 3 × 3, 5 × 5, and 7 × 7 kernels to capture complementary spatial contexts. The resulting representation is further processed by RB before classification. Comprehensive experiments were conducted on the HER2-IHC-40× dataset across three prespecified stochastic runs. Results: The mean test accuracy was 96.10% ± 0.22% and the macro-averaged F1-score was 95.30% ± 0.23%. Conclusion: Because the public release does not provide patient or whole-slide identifiers, patient-level and slide-level independence of the partitions could not be independently verified. The reported results should therefore be interpreted as an exploratory patch-level benchmark rather than as patient-level or clinically validated HER2 scoring performance.