EPDTNet + -EM: Advanced Transfer Learning and Subnet Architecture for Medical Image Diagnosis
Academic Background
In today’s healthcare environment, medical imaging plays a crucial role in disease diagnosis, treatment planning, and health management. However, traditional medical image analysis methods face numerous challenges, such as overfitting, high computational costs, limited generalization capabilities, and issues related to noise, size, and shape variations. These challenges restrict the accuracy of medical image classification and detection, thereby affecting the precision and efficiency of clinical decision-making.
To address these challenges, researchers have proposed various machine learning and deep learning-based medical image analysis methods. However, these methods still exhibit limitations when dealing with complex datasets, particularly in terms of computational efficiency and classification accuracy. Therefore, this paper introduces a novel medical image processing framework called EPDTNet+-EM (Efficient Parallel Deep Transfer Subnet + Explainable Model), which aims to improve the detection and classification accuracy of abnormalities in medical images through enhanced transfer learning and a parallel subnet architecture.
Source of the Paper
This paper is co-authored by Dhivya K, Sangamithrai K, Indra Priyadharshini S, and Vedaraj M, affiliated with SRM Institute of Science and Technology, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Vellore Institute of Technology, and R.M.D. Engineering College in India, respectively. The paper was accepted on March 13, 2025, and published in the journal Cognitive Computation with the DOI 10.1007/s12559-025-10446-w.
Research Process
1. Data Collection and Preprocessing
The study begins by collecting medical image data from multiple datasets, including brain tumor MRI, chest X-rays, chest CT scans, and mammography images. These data were obtained from various medical devices and cover a range of disease types. To ensure image quality, the research team preprocessed the raw images, including resizing, noise removal, and contrast enhancement. The preprocessing steps aim to reduce noise and distortions in the images, thereby improving the accuracy of subsequent analyses.
2. Enhanced Transfer Learning Model (EN-ETL)
After preprocessing, the research team proposed an Enhanced Transfer Learning Model (EN-ETL) for training the framework. This model employs the Enhanced Exponential Linear Unit (EN-ELU) as the activation function, replacing the traditional ReLU function. EN-ELU accelerates learning speed, enhances classification accuracy, and mitigates the “dying neuron” issue. Additionally, the model incorporates Batch Normalization and Cross-Channel Normalization to further optimize the training process.
3. Parallel Subnet Model (PSNet+)
Upon completion of training, the research team used the Parallel Subnet Model (PSNet+) to classify medical images. The PSNet+ model includes parallel convolutional layers and an axis attention mechanism, effectively balancing computational efficiency and classification performance. The parallel convolutional layers reduce computational costs, improving the model’s efficiency, while the axis attention mechanism reduces computational complexity by modeling long-range dependencies.
4. Explainable Artificial Intelligence (XAI)
To enhance the interpretability of the classification results, the research team integrated an Explainable Artificial Intelligence (XAI) model into the classification layer. XAI generates heatmaps to highlight key features in medical images, helping clinicians better understand the model’s decision-making process. This transparent explanation mechanism improves the reliability and accuracy of clinical decisions.
Key Results
1. Classification Accuracy
Experimental results show that the EPDTNet+-EM model achieved a classification accuracy of 98.83% across multiple datasets, significantly outperforming existing methods. Additionally, the model’s false positive rate (FPR) was 2%, and the false negative rate (FNR) was 3.4%, demonstrating high classification accuracy.
2. Computational Efficiency
In terms of computational efficiency, the EPDTNet+-EM model’s execution time was 5.3 milliseconds, significantly lower than other comparative methods. This indicates that the model can effectively reduce computational resource consumption while maintaining high accuracy.
3. Interpretability
Through the XAI model, the research team was able to generate detailed heatmaps, visually showcasing abnormal regions in medical images. This interpretability not only enhances the model’s transparency but also provides clinicians with more reliable diagnostic evidence.
Conclusion and Significance
The EPDTNet+-EM model successfully addresses multiple challenges in medical image classification by combining enhanced transfer learning, a parallel subnet architecture, and explainable artificial intelligence. The model excels in both classification accuracy and computational efficiency while enhancing the interpretability of clinical decisions through the XAI model. These achievements provide new solutions for the field of medical image analysis, holding significant scientific and practical value.
Research Highlights
- Enhanced Transfer Learning Model (EN-ETL): By introducing the EN-ELU activation function, the model accelerates learning speed and improves classification accuracy.
- Parallel Subnet Model (PSNet+): Through parallel convolutional layers and an axis attention mechanism, the model effectively balances computational efficiency and classification performance.
- Explainable Artificial Intelligence (XAI): By generating heatmaps, the model enhances the interpretability of classification results, providing clinicians with more reliable diagnostic evidence.
- High Classification Accuracy and Low Computational Cost: The model achieves a classification accuracy of 98.83% across multiple datasets, with an execution time of only 5.3 milliseconds, demonstrating outstanding performance.
Other Valuable Information
The research team also conducted an Ablation Study to validate the contributions of each component in the model. The results show that each component plays a crucial role in improving the model’s performance. Additionally, the team performed Cross-Validation, further verifying the model’s robustness and reliability.
The EPDTNet+-EM model provides an efficient, accurate, and interpretable solution for the field of medical image analysis, with broad application prospects.