A Hybrid Lung and Colon Histopathological Image Classification Framework Using MobileNetV3-Small Deep Features and...

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Cancer, particularly lung and colon cancer, is a major global health concern.
• Traditional histopathological diagnosis is labor-intensive, requires specialized expertise, and is prone to errors.
• Automated methods are needed to improve the efficiency and accuracy of cancer diagnosis.

Purpose of the Study:

• To develop an automated classification framework for lung and colon cancer detection using histopathological images.
• To leverage a lightweight deep learning model (MobileNetV3-Small) for efficient feature extraction.
• To optimize extracted features for enhanced classification performance and computational speed.

Main Methods:

• Utilized transfer learning with the MobileNetV3-Small deep learning model on an enhanced LC25000 dataset.
• Resized histopathological images to 224 × 224 × 3 pixels for model compatibility.
• Extracted deep features from the dropout layer and optimized them using a differential evolution algorithm, reducing dimensionality.
• Evaluated optimized features using various classifiers, including Quadratic Support Vector Machine (SVM) and bagged trees.

Main Results:

• Achieved a maximum classification accuracy of 98.14% with a Quadratic SVM.
• Demonstrated a 21.3× speed-up in classification using bagged trees.
• Outperformed several state-of-the-art methods, showing a 3.34% improvement over the baseline on the enhanced dataset.

Conclusions:

• The proposed framework effectively balances high diagnostic accuracy with computational efficiency.
• The combination of a lightweight deep learning model and feature optimization is suitable for practical clinical settings.
• This automated approach offers a promising tool for improving cancer diagnosis workflows.