MobileDANet integrating transfer learning and dynamic attention for classifying multi target histopathology images...

Area of Science:

• Medical image analysis
• Artificial intelligence in oncology
• Computational pathology

Background:

• Accurate cancer grading from histopathology is crucial for effective treatment.
• Automated computer-aided diagnosis (CAD) systems are needed to meet clinical demands.
• Deep learning (DL) offers promising solutions for complex image classification tasks.

Purpose of the Study:

• To develop and evaluate a DL framework, MobileDANet, for classifying renal cell carcinoma (RCC) into five grades.
• To extend the framework's utility to histopathology images of breast and colon cancer.
• To enhance diagnostic accuracy and efficiency in cancer severity classification.

Main Methods:

• Proposed MobileDANet architecture integrating MobileNetV2 backbone with a dynamic attention (DA) block.
• Utilized multi-head attention and MLP within the DA block to capture dependencies.
• Employed Grad-CAM for model interpretability and visualization.

Main Results:

• Achieved 90.71% accuracy and 90.94% F1 score on RCC (KMC dataset).
• Attained 88.16% recognition rate on BreakHis (breast cancer) dataset.
• Reached a 99.08% weighted F1-score on CRCH (colon cancer) dataset, outperforming baselines.

Conclusions:

• MobileDANet demonstrates high performance in classifying cancer severity across multiple histopathological datasets.
• The framework shows potential for improving automated cancer diagnosis.
• Future work includes validation on larger datasets and integration with clinical decision support systems.