Hybrid CNN-ConvFormer Model: Using Deep Learning to Improve Detection Accuracy of Histological Images for Hepatic Steatosis

This project proposes a hybrid deep learning framework combining Convolutional Neural Networks (CNN) and ConvFormer architecture, aiming to address the needs of local feature extraction and global context modeling in the detection of histological images for hepatic steatosis, thereby improving diagnostic accuracy. The project is open-sourced on GitHub, providing complete data processing and model training workflows, which have important reference value for medical image analysis and Computer-Aided Diagnosis (CAD).

Hepatic steatosis (fatty liver) is a common pathological change characterized by abnormal fat accumulation in hepatocytes. Accurate diagnosis is crucial for disease assessment and treatment. Traditional manual pathological examination is time-consuming and subjective, while deep learning-driven automated diagnosis has become a hot topic, but it faces challenges such as complex tissue structure, staining differences, and the need to capture both local details and global context simultaneously.

This model integrates the complementary advantages of CNN and ConvFormer: CNN excels at extracting local features (such as fat droplets and cell boundaries), while ConvFormer models global dependencies (such as the overall structure of normal and diseased regions) through self-attention mechanisms. The working mechanism is as follows: CNN extracts multi-scale local features, then the ConvFormer module integrates global information, simulating the diagnostic process of pathologists who first observe locally and then make an overall judgment.

The data is sourced from the Mendeley histological image dataset. The processing workflow includes: 1. Data validation (format/size check, duplicate image removal, tissue coverage evaluation, fat vacuole identification); 2. Staining standardization (unifying image staining references to improve model generalization ability). After validation, the data quality distribution is transparently reported to facilitate reproducibility.

Training uses techniques such as data augmentation, learning rate scheduling, and regularization to improve generalization ability (see requirements.txt for hyperparameters). Evaluation metrics include accuracy, precision, recall, and F1 score, among which recall (sensitivity) is particularly critical due to the high cost of missed diagnosis in medical scenarios.

The project aims to develop a CAD system for auxiliary pathological diagnosis, serving as a "second opinion" to improve diagnostic consistency and efficiency; it can be used for quantitative analysis of fat infiltration percentage to support disease monitoring and treatment evaluation; it also provides a tool foundation for medical researchers, which can be extended to automated diagnosis of other liver diseases.

This hybrid model represents the cutting-edge direction of medical image analysis, and the integration of local and global features demonstrates the potential of deep learning in medical diagnosis. The project provides a complete workflow reference, which is valuable for AI researchers, developers, and clinicians. In the future, similar systems are expected to reduce the burden on doctors in clinical practice and benefit patients. Its architecture and technical solutions (such as staining standardization) can be migrated to other medical image tasks and multi-organ pathological analysis.