Alibaba Open-Sources RADAR: One Model Detects 146 Diseases

Alibaba Advances Medical Imaging AI from “One Model per Disease” to “One Model for Multiple Diseases”

On September 18, Alibaba’s DAMO Academy announced the open-source release of the general-purpose medical imaging model DAMO RADAR. Jointly developed by DAMO Academy, the First Affiliated Hospital of Zhejiang University School of Medicine, and other institutions, the model is designed for contrast-enhanced abdominal CT and can identify more than 146 conditions in a single examination. The related findings have been published in Science.

According to the publicly available results, DAMO RADAR achieved an average AUC of 0.913 across nearly 40,000 real-world examinations. In a human–AI comparison study, its average accuracy exceeded that of 23 of the 26 participating radiologists. When provided with model-generated prompts, physicians improved their disease-detection sensitivity by 10% and reduced image-reading time by more than 30%.

This is not just another disease-specific model that can detect only pulmonary nodules or pancreatic cancer. DAMO RADAR attempts to address a long-standing structural problem in medical imaging AI: hospitals operate in a real world containing hundreds of abnormalities, while most previous algorithms have had to be trained and deployed separately for each disease.

146 Conditions: What Really Matters Is Detecting Them Simultaneously

Abdominal CT is one of the most complex tasks in radiology.

A contrast-enhanced CT examination typically contains hundreds of slices covering multiple tissues and organs, including the liver, gallbladder, pancreas, spleen, gastrointestinal tract, kidneys, and blood vessels. Physicians cannot focus solely on the area associated with the patient’s chief complaint. They must also check for tumors, inflammation, obstruction, hemorrhage, vascular abnormalities, and incidental findings.

Traditional medical AI generally follows a “one model per disease” approach:

• One model for liver cancer;
• Another model trained for pancreatic cancer;
• A separate model for aortic dissection;
• Switching to another hospital or scanner may require the model to be adapted all over again.

This approach can easily achieve high scores in research papers and single-disease screening tasks, but it quickly encounters engineering challenges when deployed in hospitals. If a hospital wants to cover dozens of diseases, it must deploy dozens of models and separately maintain their preprocessing pipelines, inference services, thresholds, alert logic, and versions. The models may also produce mutually conflicting results.

DAMO RADAR’s direction is closer to how large language models have replaced numerous single-task NLP classifiers: instead of building a new tool from scratch for every disease, it first trains a foundation model capable of understanding relationships across multiple organs and diseases, and then uses that model to perform multilabel recognition in a single imaging examination.

It is important to note that “identifying more than 146 conditions in a single examination” does not mean the model can make a definitive clinical diagnosis from a single CT image. A more accurate interpretation is that the model can simultaneously screen a complete contrast-enhanced abdominal CT study for multiple diseases and output probabilities or alerts for different abnormalities, which radiologists can then review. A final diagnosis must still take into account the patient’s medical history, laboratory tests, pathology results, and clinical presentation.

The Technical Key: More Than Simply Training on CT Scans and Reports Together

DAMO RADAR uses a vision–language learning method that enables the model to learn associations between CT images and the corresponding radiology reports.

The concept appears similar to image–text models such as CLIP: images are provided on one side and text on the other, with contrastive learning used to bring images and reports expressing the same content closer together. However, medical CT is far more difficult than natural-image processing.

The subject of an ordinary photograph is usually clear and centrally located. In a photograph of a cat, for example, the cat generally occupies a large portion of the image. By contrast, an abnormality in an abdominal CT study may appear in only a few of hundreds of slices, or may consist of nothing more than a low-density region a few millimeters wide within an organ. Directly aligning an entire CT study with a complete report is like asking a model to find the punctuation mark corresponding to a particular sentence in a book hundreds of pages long—the signal can easily be drowned out.

DAMO Academy’s solution is fine-grained organ-level alignment.

The model first constructs the CT scan as three-dimensional data and then decomposes it into units representing different organs or tissues based on anatomical structure. This allows image regions to correspond at the organ level with relevant descriptions in the report. For example, a report’s description of a pancreatic mass should primarily be associated with the pancreatic region rather than broadly matched with the entire set of abdominal images.

This method primarily addresses three problems:

1. Reducing noise in three-dimensional CT scans. The model does not need to search blindly for lesions across the entire abdominal cavity; instead, it can learn features associated with specific organs.
2. Using existing reports as supervisory signals. The large volumes of CT scans and radiology reports accumulated by hospitals can be used for training, eliminating the need to create new pixel-level or bounding-box annotations for every disease.
3. Improving the interpretability of results. When the model flags a disease, it can link its assessment to the corresponding organ, image region, and textual semantics rather than outputting only an untraceable score.

The research team also uses an adaptive contrastive modeling strategy to dynamically adjust training. This matters because different diseases vary greatly in prevalence, imaging appearance, and reporting terminology. The model cannot treat common diseases, rare diseases, and normal findings as samples of equal difficulty.

This is more valuable than simply increasing the model’s parameter count. The bottleneck in medical imaging has never been computing power alone; it is also the granularity of the data. Images are three-dimensional, lesions are localized, reports are unstructured, and disease distributions are extremely imbalanced. Without a more fine-grained alignment mechanism, increasing the volume of data does not necessarily enable the model to learn more reliable medical features.

An AUC of 0.913: How Good Is It?

Across the 146 conditions evaluated by the research team, DAMO RADAR covers 18 organs and achieved an average AUC of 0.913 in nearly 40,000 real-world examinations.

AUC measures a model’s ability to rank positive cases ahead of negative cases. A value of 0.5 is close to random guessing, while 1 represents perfect discrimination under ideal conditions. A score of 0.913 indicates that the model has strong overall discriminative capability, but it does not mean that “91.3% of patients can be diagnosed correctly,” nor can it directly substitute for sensitivity and specificity at a particular clinical threshold.

The human–AI comparison results reported in the paper are more representative of real-world use:

• DAMO RADAR’s average accuracy exceeded that of 23 of the 26 participating physicians;
• With AI-generated prompts, physicians improved their disease-detection sensitivity by 10%;
• Image-reading time was reduced by more than 30%;
• With AI assistance, junior physicians reached the diagnostic performance of senior physicians.

These results suggest that DAMO RADAR’s most realistic current role is not to “replace radiologists with AI,” but to serve as a second reader with broader coverage.

When reviewing scans, physicians often worry less about being unable to assess a typical lesion they have already found than about overlooking an abnormality unrelated to the patient’s chief complaint during an intense workload. A general-purpose model can systematically scan 18 organs, flag potential issues, and then leave confirmation to the physician. This is similar to a code review tool: it does not make the final decision for the developer, but it can quickly identify risks that might otherwise be overlooked.

The study also reported that DAMO RADAR achieved an AUC of 0.904 on acute abdominal conditions outside its original training scenario. This result is noteworthy because emergency department data is typically more chaotic: disease distributions differ, patients have more complex conditions, and scanning protocols may not be entirely consistent. The model’s ability to maintain similar performance in such settings provides preliminary evidence of cross-scenario generalization.

However, a high AUC in a single study does not mean the model is ready for direct use in every hospital. The most common problems with medical AI arise precisely when the data distribution changes, for example:

• Imaging differences caused by CT scanners and reconstruction algorithms from different vendors;
• Inconsistencies in contrast-enhancement phases, slice thicknesses, and examination workflows across hospitals;
• Underrepresentation of children, older adults, or rare cases in the training data;
• Differences in report templates, disease terminology, and physicians’ writing styles;
• Potentially severe sample-size imbalances among the 146 diseases.

The average AUC is therefore only a starting point. Developers and medical institutions must also examine performance for each individual disease, confidence intervals, results on external test sets, and false-negative and false-positive rates at specific thresholds. A model that performs extremely well on common diseases may conceal inadequate detection of rare but fatal conditions.

Open Source Matters, but It Is Still Far from “Download and Use”

DAMO RADAR is now publicly available on GitHub. For medical AI, the value of open source lies not only in allowing developers to run the model, but more importantly in enabling research institutions to inspect the training paradigm, reproduce experiments, and conduct external validation using data from different hospitals.

Many previous medical AI studies offered only papers and closed demonstrations, making it difficult for third parties to determine whether the results came from genuine generalization or meticulous tuning on a specific dataset. Opening the code and related resources at least provides a foundation for further iteration on the fine-grained organ-level alignment approach.

However, open source does not mean that hospitals can directly connect the repository to their PACS systems. Real-world deployment still requires work at several levels:

1. Data Pipeline

Raw hospital data typically consists of DICOM series rather than model-ready tensors. Engineering teams must handle series selection, phase identification, voxel-spacing standardization, resampling, organ localization, and anomaly checking. An error at any stage could cause the model’s input to deviate from the configuration used in the paper.

2. Three-Dimensional Inference Costs

Abdominal CT studies contain large numbers of slices, and three-dimensional models have significantly higher memory and inference costs than conventional two-dimensional image classifiers. Hospitals must also account for concurrent examination volumes, emergency prioritization, inference latency, and GPU utilization—not merely whether the model can process a single sample.

3. Clinical Workflow

The model’s output cannot be limited to a JSON file or a list of probabilities. It must be integrated into the image-viewing interface physicians already use, linked to specific series and suspicious regions, and accompanied by records of the model version, input information, thresholds, and operation logs. Otherwise, even a highly accurate algorithm will struggle to deliver meaningful reductions in reading time.

4. Regulatory and Liability Boundaries

Research models, computer-aided diagnostic software, and medical devices capable of independently producing diagnostic conclusions represent entirely different product stages. Open-source code can be used for research and validation, but clinical deployment still involves regulatory approval, privacy protection, security assessments, and continuous monitoring.

Before using the repository, developers should also review the exact materials provided, the license, how model weights can be obtained, environment dependencies, and the permitted scope of use. Medical data is inherently highly sensitive, and open-sourcing the model does not eliminate data-compliance requirements.

The Value of This Release Is Not That “AI Beat 23 Physicians”

The claim that its “average accuracy exceeded that of 23 physicians” is the most eye-catching result, but it is not the most important aspect of DAMO RADAR.

Physician comparison studies are influenced by case selection, reading time, participant experience, and evaluation criteria. Their results cannot simply be interpreted as evidence that AI has comprehensively surpassed radiology experts. Medical diagnosis is also not a single leaderboard: physicians can ask about medical history, review previous examinations, incorporate laboratory indicators, and assume ultimate responsibility for ambiguous results. The model currently cannot do these things.

DAMO RADAR’s more substantive breakthrough is demonstrating that a single model can cover 18 organs and more than 100 diseases in abdominal CT, while obtaining scalable supervisory signals from naturally accumulated image–report data.

This could change the cost structure of medical AI.

Previously, adding a disease often meant collecting new cases, asking physicians to annotate them, training a new model, and establishing a new workflow. With a general-purpose model, adding new tasks could instead become a matter of adaptation and validation on top of a unified representation. For hospitals, this is far more practical than deploying more than 100 mutually incompatible disease-specific models.

Of course, DAMO RADAR still focuses exclusively on contrast-enhanced abdominal CT. There is a long way to go before truly general-purpose medical models can cover multimodal medical data such as chest imaging, cranial imaging, MRI, ultrasound, and pathology. Different imaging modalities have vastly different physical mechanisms, spatial resolutions, and diagnostic logic; these differences cannot be addressed simply by adding more disease labels.

But the direction is now clear: medical imaging AI is evolving from narrowly focused point solutions toward foundation models capable of understanding anatomical structures, disease semantics, and clinical reports.

For developers, DAMO RADAR is not a new model suited to building chatbots, nor is it a capability that can be directly replaced by an ordinary cloud API. It is more like an infrastructure blueprint for three-dimensional medical data: how to use weakly labeled reports, how to perform organ-level alignment, how to train on long-tailed distributions across multiple diseases, and how to embed a model into high-risk clinical workflows.

If prospective validation at external hospitals continues to produce results close to those currently reported, DAMO RADAR’s significance will extend beyond adding another paper to a top-tier journal. It could make it possible, for the first time, for medical imaging AI to evolve from a “specialized detector” into a “general-purpose image-reading assistant.”

References

• ITHome: Alibaba DAMO Academy Open-Sources DAMO RADAR, Capable of Identifying More Than 146 Conditions — Introduces the model release, technical approach, human–AI comparison study, and key performance data.
• GitHub: DAMO RADAR Open-Source Repository — The project’s official open-source repository, where subsequently released code, model resources, and usage instructions can be found.