A Hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM)-Attention Model Architecture for Precise...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning (DL) models, particularly Convolutional Neural Networks (CNNs), are increasingly vital for accurate medical image classification.
• Existing methods require enhancement for improved diagnostic speed and precision.

Purpose of the Study:

• To introduce MediVision, a hybrid deep learning model designed to advance medical image classification.
• To improve the accuracy and interpretability of automated disease diagnosis.

Main Methods:

• MediVision integrates a CNN backbone for feature extraction, Long Short-Term Memory (LSTM) for sequential dependency identification, and an attention mechanism for salient feature focus.
• A skip connection and Grad-CAM heatmap are employed to enhance feature representation and visualize critical image regions.

Main Results:

• MediVision demonstrated consistent classification accuracy exceeding 95% across ten diverse medical image datasets.
• The model achieved a peak accuracy of 98% in classifying conditions such as Alzheimer's disease, various cancers, and retinal diseases.

Conclusions:

• MediVision provides a robust framework for reliable and interpretable medical image classification.
• The study enhances automated disease diagnosis and supports research reproducibility through open-access code and datasets.