Micro-morphological feature visualization, auto-classification, and evolution quantitative analysis of tumors by...

Area of Science:

• Medical Imaging
• Pathology
• Artificial Intelligence

Background:

• Accurate tumor visualization and quantitative analysis are crucial for early cancer detection and understanding metastasis.
• Current medical imaging techniques face challenges in multiscale, 3D, non-destructive pathological assessment.

Purpose of the Study:

• To develop and validate a novel method for high-resolution 3D pathological visualization and quantitative analysis of digestive system tumors.
• To apply deep learning for automated classification of tumor types based on micro-morphological features.

Main Methods:

• Utilized synchrotron-based X-ray phase-contrast tomography (SR-PCT) with phase-and-attenuation duality phase retrieval for 3D tumor reconstruction.
• Extracted a feature set of eight tumor micro-lesion types from high-density resolution SR-PCT data.
• Trained an AlexNet-based deep convolutional neural network for automated tumor classification.
• Employed machine learning methods (AUC, PCA) to analyze micro-pathomorphological relationships in liver tumors.

Main Results:

• Achieved 94.21% average accuracy in auto-classifying eight types of digestive system tumors using the deep learning model.
• Revealed micro-pathomorphological relationships of liver tumor angiogenesis and progression through quantitative feature analysis.
• Demonstrated that tumor lesion progression is linked to the inflammation microenvironment.

Conclusions:

• High phase-contrast 3D pathological characteristics derived from SR-PCT offer excellent recognizability and classifiability for micro tumor lesions.
• The developed automatic analysis methods show significant potential for improving tumor typing and statistical calculations.
• This approach aids in understanding tumor evolution and its clinical manifestations.