From detection to reading
With M4CXR, our proprietary medical-imaging foundation model, DEEPNOID is creating a new category of chest X-ray AI.
M4CXR is a medical-imaging foundation model pre-trained on more than 10 million clinical records from tertiary hospitals.
Built on a multimodal architecture that learns images and reports together, it expands to new findings and tasks without repeated development,
to lift chest X-ray AI from ‘single-finding detection’ to ‘clinical reading.’
Repeated Development, Rising Costs
Conventional task-specific deep-learning models must repeat the entire cycle—
data collection, labeling, model design, training, and validation—for every new finding.
Each added finding effectively costs as much time and effort as building a brand-new model,
which is why real clinical settings could only offer limited functionality.
M4CXR is designed to resolve this structural bottleneck.
Learning That Starts From Radiology Reports
M4CXR learns clinical radiology reports as natural language, as written.
That means it scales training and expands quickly without labeling cost.
Natural-Language Supervision
M4CXR learns from the radiology report data generated every day in clinical practice, without additional labeling. This gives it a structural advantage: training can keep scaling without rising labeling costs.
Rapid Task Expansion
M4CXR does not build a new model from scratch for each new finding. It expands with minimal fine-tuning based on pre-trained representations, fundamentally resolving the ‘new model per finding’ cost structure described above.
Scalability Solved With a Single Model
Across all four axes—data, architecture, scalability, and performance—
M4CXR takes a different approach from task-specific models.
| Traditional AI | Dimension | M4CXR |
|---|---|---|
| Limited to a single data type: image or text | Data | Pre-trained on multimodal data such as images and text |
| Task-specific models designed per individual finding | Architecture | Generally applicable across tasks, findings, and clinical settings |
| Rebuilt from scratch for each new finding | Scalability | Reusable structure and pre-training accelerate new development |
| Struggles to adapt to changes in data types and clinical environments | Accuracy | Higher accuracy and broader finding coverage vs. single-modality models |
Performance Proven in Clinical Settings
Trained on more than 10 million clinical records,
M4CXR reads 41+ findings in a single inference within 2.3 seconds.
Clinical data used for training
Findings handled in a single inference
Seconds to complete a single inference
Our Proprietary-Model Strategy
To reduce the licensing and dependency risks of external general-purpose models,
M4CXR was designed as a proprietary foundation model specialized for medical imaging.
General-purpose foundation models carry structural risks:
licensing-policy changes, data-usage limits, and external tech dependency.
Designed as a proprietary model specialized for medical imaging,
M4CXR secures technical independence, data security, and long-term cost efficiency at once.
License Independence
Independent development, free from external constraints
Data Security
Safe use of medical data
Continuous Improvement
Domain-specialized performance upgrades
Cost Optimization
Efficient operations and maintenance
From a Single Imaging Model to a Clinical Agent
Starting from chest X-ray, M4CXR spans diverse medical imaging
and ultimately extends into a medical AI agent that performs clinical decision-making.
M4CXR is a foundation model proven in the chest X-ray domain.
The next step, MedZero-27B, is a 27B-scale expansion model
that learns 2D and 3D medical images—X-ray, CT, MRI, and more—together,
designed to handle diverse clinical domains within a single model rather than one imaging type.
The final step, Clinical Agent, goes beyond image reading:
it autonomously integrates and reasons over patient information, prior exam history, and clinical guidelines,
actively performing reading, finding summarization, and follow-up exam recommendations.