MD.ai

MD.ai works as a radiology annotation tool and a viewer on the same data. The platform handles DICOM from the ground up rather than as an import format. You add exams by direct upload, by connecting a cloud bucket on Google Cloud, Amazon S3, or Azure, or by pushing them over DICOM C-STORE. JPEG, PNG, TIFF, and video files can also come in through a patient-centric folder layout, which matters when a study mixes imaging and clinical clips.

The viewer carries FDA 510(k) clearance, so clinical interpretation, review, and reporting happen in a cleared tool rather than a research viewer with a disclaimer. It supports standard zoom, pan, and windowing, hanging protocols, multiplanar reconstruction, and measurement tools. Annotations sit in the same workspace as the image, so you label and read in one screen.

You can deploy your own model and run distributed inference on the dataset. Output arrives as a pre-annotation that the annotator accepts, adjusts, or rejects. Models can be deployed per project with their own settings, and the platform records model performance against the human labels, which is the feedback loop most annotation tools leave out.

Federated annotation lets several institutions join one AI model validation run without pooling patient data. Each site labels or runs inference locally, and only the results travel. For a vendor that needs external validation before a submission, this is often the only way to get multi-site evidence without months of data-sharing agreements.

Built-in PHI detection scans images and metadata for identifying information and strips it before export. Pixel-level de-identification handles burned-in text where it appears on the image itself. It's the step that decides whether a dataset can leave a hospital at all, and having it in the platform saves a separate compliance project.

As radiology reporting software, MD.ai uses language models for the tasks that eat a radiologist's day: picking the right template, mapping dictated findings into the right fields, drafting the impression, proofreading, and generating billing codes at signature. You can run it as a guided, AI-driven mode or stick to a traditional dictation mode and call the assistance when you want it.

Reports can pull in guideline logic for BI-RADS, LI-RADS, TI-RADS, and other systems, and you can add your own custom guidelines. The system also compares the new study with prior reports and flags what changed. Both features cut the back-and-forth that makes a report take longer than the reading itself.

A CLI tool and a Python client library let teams create projects, upload data, pull annotations, and manage users from code. Reporting connects to EHR, HIS, and RIS through HL7 and DICOM, and can be embedded in another application with an iframe using token-based access, so the report opens inside the hospital's own workflow.