End to End Multitask Joint Learning Model for Osteoporosis Classification in CT Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Osteoporosis is a prevalent global health issue often detected late due to asymptomatic progression.
• Current diagnostic methods like dual-energy X-ray and CT scans are costly and time-consuming.
• Developing efficient and economical osteoporosis diagnosis tools is crucial.

Purpose of the Study:

• To propose a joint deep learning framework for enhanced osteoporosis diagnosis.
• To integrate localization, segmentation, and classification for improved accuracy.
• To address the limitations of existing methods requiring time-consuming lesion annotation.

Main Methods:

• A novel joint learning framework combining localization, segmentation, and classification for osteoporosis diagnosis.
• Implementation of a boundary heat map regression branch for segmentation and a gated convolution module for classification.
• Integration of segmentation and classification features using a dedicated feature fusion module.

Main Results:

• The proposed model achieved an overall accuracy of 93.3% on a self-built dataset for classifying normal, osteopenia, and osteoporosis.
• High Area Under the Curve (AUC) scores were obtained: 0.973 for normal, 0.965 for osteopenia, and 0.985 for osteoporosis.
• The feature fusion module effectively adjusted the weight of different vertebral levels for improved diagnostic performance.

Conclusions:

• The developed joint learning framework offers a promising and accurate approach for osteoporosis diagnosis.
• This method presents a more efficient and economical alternative to current diagnostic techniques.
• The study highlights the potential of deep learning in improving early detection and management of osteoporosis.