DeepRad.AI’s four major disease modules predict the risk of osteoporosis, lung cancer, COPD, and coronary artery calcification in a single assessment.

Originally published in New Venture

The Artificial Intelligence Medical Research Center at Taipei Medical University, led by Vice President Dr. Cheng-Yu Chen, has developed the AI-powered multi-modal imaging precision health platform “Deep-Lung.” With just a single low-dose chest CT scan, the proprietary AI model can simultaneously predict the risks of osteoporosis, lung cancer, chronic obstructive pulmonary disease (COPD), and coronary artery calcification, achieving an accuracy rate of over 90%. The system also automatically generates recommendation reports that meet international standards, making it the first of its kind in the world.

Dr. Yen-Ting Chen, who oversees platform integration, stated that during hospital validation, the platform successfully identified lung nodules that clinical physicians had missed or failed to annotate multiple times…

Taipei Medical University’s AI medical team is set to launch another startup, DeepRad.AI, to further advance the application of its groundbreaking software platform, Deep-Lung—the world’s first AI-driven multi-modal imaging precision health platform capable of early screening for four diseases affecting the heart, lungs, and bones using a single low-dose lung CT scan.

Deep-Lung currently features four major modules:

Lung Nodule Module (LungRads)

The system is an automated lung nodule detection and interpretation platform, comprising a nodule detection model, a nodule segmentation model, a benign-malignant classification model, and an automated Lung-RADS reporting model. It provides multi-nodule localization with a sensitivity of 94%, accurate nodule size measurement with a Dice coefficient of 81.39%, and lung nodule benign-malignant classification with an accuracy of approximately 98%.

The final module automatically generates a comprehensive LungRads text report, incorporating results from the previous analyses. This report includes detailed information on lung nodule texture, shape, size, location, and classification according to international lung nodule categorization and management standards, making it the first of its kind in Taiwan.

This module was trained using data from nearly 6,000 cases collected through the Ministry of Science and Technology’s Big Data Imaging Project, sourced from Taipei Medical University Hospital, Shuang Ho Hospital, and Wan Fang Hospital. Each lung nodule was meticulously annotated by specialist physicians, with annotations covering 107 radiomic features, 47 semantic labels, and corresponding pathology reports to ensure precise and reliable analysis.

※ The image shows the Deep-Lung AI platform interface. (Photo courtesy of Taipei Medical University)

Coronary Artery Calcification Module (CAC)

This module enables the prediction of the Coronary Artery Calcification (CAC) score directly from low-dose chest CT scans, achieving results nearly equivalent to those obtained from standard high-dose coronary CT scans. It provides cardiovascular risk stratification and health recommendations based on American Heart Association (AHA) guidelines.

Clinical validation using 202 cases from Taipei Medical University Hospital and 549 cases from Shuang Ho Hospital demonstrated that the module achieves a sensitivity of 92% and 92.7%, respectively, in detecting high-risk patients (CAC score > 400).

Additionally, a key advantage of this module is its superior AI training dataset compared to global literature. Each low-dose chest CT scan is paired with a same-day high-dose coronary artery CT scan to obtain the reference CAC score. This dataset, derived from 1,621 cases at Taipei Medical University Hospital, serves as the gold standard for correlation assessment, ensuring higher accuracy and reliability.

Emphysema Module (COPD)

This module utilizes low-dose chest CT images to precisely segment lung lobes and automatically calculate the Emphysema Severity Index (RA950: percentage of the relative area of the lungs with attenuation values < -950 Hounsfield units). It then predicts the likelihood and severity of chronic obstructive pulmonary disease (COPD).

Clinical validation shows that the current lung lobe segmentation accuracy achieves a Dice coefficient of 95.6%. Additionally, the RA950 index automatically derived from chest CT scans exhibits a strong correlation with patients’ lung function (regression coefficient -0.72, p < .00001). The module also achieves a 90.2% sensitivity (ROC-AUC = 0.813) in screening for COPD patients.

Bone Mineral Density Prediction Model (BMD)

Multiple studies have shown that thoracic vertebral bone density can be used to accurately predict the degree of osteoporosis. This model is one of the few innovative approaches that directly predicts bone mineral density from thoracic spine CT images. It provides bone density classification, including normal, low bone mass, or osteoporosis.

The ground truth for this model’s training data is based on standard DXA bone mineral density measurements of the lumbar spine and hip, taken on the same day as the subject’s low-dose chest CT scan. As a result, the prediction reliability is extremely high. After testing at Taipei Medical University Hospital and external institutions, the model has achieved accuracy rates of 88% and 94%, with sensitivity rates of 89% and 94%, respectively.

Yen-Ting Chen, Director of the Taipei Medical University Artificial Intelligence in Medicine Research Center, who is responsible for the development and integration of Deep-Lung, stated that the team will continue to collect a large volume of clinical case data to train the model, ensuring that the detection and interpretation results provided by the platform become more accurate and closer to clinical practice.

He also mentioned that while research in the past was primarily conducted for academic publication, the current goal is to translate research into real-world applications. One of the most meaningful aspects of this platform, in his view, is its ability to automatically generate reports that meet international standards, saving clinicians valuable time. Additionally, during clinical validation, the platform has occasionally detected lung nodules that physicians either failed to annotate or overlooked, highlighting one of its most significant contributions and values.

The commercial applications of Deep-Lung will include domestic and international sales and licensing, a personal health cloud platform, risk prediction for insured individuals, new insurance policy development, and new drug trials. The team has identified its primary target customers as hospital radiology departments, health screening centers, and health consultation centers, while secondary and potential customers include middle-aged and older adults over 50, the insurance industry, and pharmaceutical companies.

The team plans to complete a regulatory clinical trial this year and will successively apply for certification and approval in Taiwan and the United States. Yen-Ting Chen stated that approvals will first be sought for individual indications, with a possible later application for the integrated four-in-one use case. Future expansion will also target markets in the EU, Australia, Japan, South Korea, and Southeast Asia, enabling middle-aged and elderly individuals to receive simultaneous lung, heart, and bone screening recommendations and health assessments with minimal CT radiation exposure, effectively reducing disease risks.