Protein Localization in Immunohistochemistry Images using Multi-scale Features and Deep Learning

Protein subcellular localization is crucial for understanding protein functions and disease mechanisms. Immunohistochemistry (IHC) images provide valuable insights into protein distribution at the tissue level, supporting localization studies. Despite advances in computational methods for IHC-based subcellular localization, challenges remain in feature representation and complexity. To address these, we propose iMFDeepLoc, a deep learning model that integrates 13 features across multiple scales, including traditional features (LBP, Haralick, HOG, Gabor), color, spatial relations, cavity features, and YOLO V8-based deep features. Feature selection is performed using stacked denoising autoencoders (SDA) and dimensionality reduction with linear discriminant analysis (LDA) to optimize the feature set. The model employs a CNN-BiLSTM-Attention architecture for efficient feature learning. Our results show that combining local, global, and deep features at multiple scales with deep learning improves predictive performance, achieving an overall accuracy of 76.2%. This approach offers a novel and effective solution for predicting protein subcellular localization, providing valuable support for biological research.