Alibaba Damo Radar: Open-Source Medical AI for Abdominal CT Diagnosis

Alibaba Group Holding’s Damo Academy has open-sourced Damo Radar, a vision-language model designed to identify nearly 150 different abdominal conditions, including various cancers, through the analysis of contrast-enhanced computed tomography (CT) scans. The model aims to serve as a generalist medical imaging tool to assist radiologists in improving diagnostic accuracy and reducing analysis time.

Diagnostic Performance and Benchmarks

Damo Radar achieved an average area under the curve (AUC) of 0.913 across 146 clinical findings when tested on nearly 40,000 real-world examinations. The model was trained using a dataset of CT scans paired with corresponding clinical reports, covering 18 different abdominal organs.

In a comparative study involving 26 human radiologists from multiple hospitals, Damo Radar's average accuracy exceeded that of 23 of the participants. When used as a supportive tool, the model provided the following benefits to clinicians:

  • Reduction in missed diagnoses: Radiologists improved their ability to prevent missed diagnoses by 10%.
  • Efficiency gains: The time required for diagnosis was reduced by more than 30%.

Technical Architecture and Methodology

The research associated with Damo Radar was published in the journal Science. The model is described as a generalist medical imaging model, and the research team suggests that the training methodology can be extended to other modalities of medical imaging beyond abdominal CT scans.

Technical discussions among practitioners highlight a specific architectural detail: the CNN-based feature encoder used in the model significantly outperformed the vision transformer (ViT) encoder backbone in this specific medical imaging application.

Context within Alibaba's Medical AI Portfolio

Damo Radar is part of a broader strategy by Alibaba to integrate AI into healthcare. This release follows the Coca AI model, released in April, which was specifically designed for the early detection of colorectal cancer from CT scans and reported higher sensitivity than human radiologists in spotting early-stage malignancies.

Community Perspectives and Critiques

While the reported AUC is high, some technical critics argue that AUC is an insufficient metric for medical diagnostics due to severe class imbalance (where the disease is far less common than the healthy state).

"ROC-AUC of 0.9 under severe class imbalance... could still mean something like 4/5 predicted diagnoses are wrong (false positives). Precision-Recall curve + mAP or GTFO."

This critique suggests that without precision-recall metrics, the true clinical utility and the rate of false positives in a real-world setting may be obscured by the AUC figure.

Sources