A deep learning ICDNET architecture for efficient classification of histopathological cancer cells using Gaussian noise images

Histopathology image classification is crucial in pathological diagnosis workflow for early detection and treatment. The integration of deep learning technology has greatly improved diagnostic accuracy and efficiency. However, there are limitations when morphological features are not obvious in pathological sections, leading to difficulties in identifying deep cells and an increased risk of misdiagnosis. To address this issue, this study introduces a new hybrid network model, termed ICDNET, designed to fuse global and local features without destroying the integrity of the feature data, thus enhancing the accuracy of medical image classification. The ICDNET model consists of two main features: (i) a serial hierarchical structure composed of global and local feature blocks; and (ii) an Internal Communication Hierarchical Fusion Block (ICHF) and an Efficient Dual Self-Attention (EDA) mechanism. This network structure solves internal communication issues and enriches contextual semantic information, extracting local features and global representations from different internal spaces. To evaluate the performance of the ICDNET network model, experiments were conducted on four major public datasets with the addition of Gaussian noise. The experimental results demonstrate excellent accuracy and the ability to handle limited training samples, highlighting the potential of the ICDNET model to assist pathologists in pathological diagnosis.