Towards an Informed CNN for Bone SR-microCT Image Classification with an Unsupervised Patched-based Image Clustering

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medicine
• Bone Biology

Background:

• Human visual inspection of Synchrotron Radiation micro-Computed Tomography (SR-microCT) images struggles to identify subtle bone microscale differences.
• Deep Learning (DL) offers potential for analyzing complex imaging data but often requires guidance to focus on relevant details.

Purpose of the Study:

• To develop and evaluate a DL-based method for classifying femoral head images from SR-microCT scans.
• To differentiate between healthy, osteoporotic, and COVID-19 affected bone tissue using microscale features.

Main Methods:

• Utilized unsupervised patch-based clustering to inform a vgg16 model.
• Focused on subtle microscale differences in SR-microCT images of femoral heads.
• Applied the method to classify images into healthy, osteoporotic, and COVID-19 categories.

Main Results:

• Achieved up to 9.8% accuracy improvement in classifying healthy versus osteoporotic images compared to uninformed methods.
• Demonstrated 59.1% accuracy in distinguishing between osteoporosis and COVID-19 affected bone.
• Established a classification accuracy of 60.91% for healthy versus osteoporotic bone.

Conclusions:

• The proposed DL method effectively classifies SR-microCT images of femoral heads based on subtle, human-imperceptible microscale differences.
• This approach provides a foundation for automated diagnosis of bone conditions like osteoporosis and COVID-19 related bone changes using advanced imaging techniques.