RADIA: Beyond DICOM AI for Medical Imaging in Fundoscopy & Ultrasound

Most medical AI platforms focus on a single niche: chest CT, digital mammography, or bone X-rays. But the reality for healthcare professionals is far more diverse. An ophthalmologist needs to analyze fundus photographs looking for signs of glaucoma or intraocular tumors. A musculoskeletal radiologist reviews Doppler ultrasounds to differentiate a hemangioma from a sarcoma. A rural physician wants a second opinion on that fundus image they just captured with a portable retinal camera.

RADIA was born as a DICOM viewer with AI for CT and dental CBCT, but its multimodal architecture — four AI models analyzing each image in cascade — works equally well with any clinical image. Today we show how it processes two modalities that go far beyond classic DICOM: fundus photography (retinography) and soft tissue ultrasound with color Doppler.

Retinography — fundus photography — is one of the most informative and least invasive tests in medicine. Through the dilated pupil, a camera captures the retina, optic disc, macula, and retinal vessels. In that single image, dozens of pathologies can be detected: from diabetic retinopathy and glaucoma to intraocular tumors and signs of intracranial hypertension.

The problem is that interpreting a fundus photograph requires specialized experience. A primary care physician may not recognize a cotton wool spot or might confuse a benign nevus with a choroidal melanoma. RADIA tackles exactly this problem: the AI analyzes each image with the same systematic approach as an experienced retinologist, identifying findings, classifying them by severity, and suggesting clinical differentials.

The first case shows a fundus photograph with a pigmented oval lesion in the macular region, temporal to the optic disc. This is the type of finding that can trigger immediate alarm — is it a benign choroidal nevus or an early melanoma?

RADIA identified two findings:

What's relevant here isn't just the detection — any trained eye sees the dark spot. What matters is the categorization. RADIA doesn't say "malignant tumor" when it isn't. It classifies the findings as indeterminate, notes the absence of orange pigment, overlying drusen, or subretinal fluid (markers of malignant transformation), and suggests a benign choroidal nevus as the most probable finding. That precision avoids unnecessary referrals and patient anxiety.

Retinoblastoma is the most common malignant intraocular tumor in childhood. Detecting it early can save not only the eye but the child's life. And when it's multicentric — multiple independent foci — the urgency multiplies.

In this case, RADIA analyzed the image with Asteroid Mode (deep multi-model analysis) and detected 5 malignant findings, all classified as Neoplasm with 98% confidence.

Each focus was described individually: "elevated intraocular lesion, bright white-yellowish and highly reflective, with a granular texture suggestive of calcification." The presence of calcification is the most specific marker of retinoblastoma on imaging, and RADIA identified it in all foci.

Papilledema is optic disc swelling caused by intracranial hypertension. It's a neurological emergency diagnosed by looking inside the eye. Brain tumors, venous sinus thrombosis, pseudotumor cerebri — all can first manifest as a blurred, elevated optic disc on fundoscopy.

RADIA described the finding with clinical precision: "markedly elevated optic disc, hyperemic with blurred and indistinct margins in all directions, especially in the superior, nasal, and inferior quadrants. The physiological cup is obliterated. Significant peripapillary edema extending into the surrounding retina."

The Deep Analysis included a complete clinical differential:

And most importantly: the red-highlighted recommendation — "Urgent neurological evaluation to rule out intracranial hypertension, including neuroimaging (contrast-enhanced brain MRI and venous MRI)." RADIA doesn't just detect, it prioritizes. A finding requiring urgent neuroimaging is flagged as such, never lost among data.

This case is the most complex in the ophthalmology set. A fundus photograph showing diffuse vascular changes, cotton wool spots, papilledema, and retinal hemorrhages — all in the same image. It's the kind of case that demands analyzing multiple interrelated findings and synthesizing a coherent clinical picture.

The AI synthesized all these findings into a coherent clinical picture: "severe and diffuse retinal microangiopathy, consistent with ischemic retinopathy." The clinical differential included severe/malignant hypertensive retinopathy as the leading option. This isn't a list of isolated findings — it's integrated clinical reasoning.

One of RADIA's most powerful features is PointAsk: the doctor clicks on any point in the image and the AI specifically analyzes that area. Not a general analysis — a focal analysis at the exact coordinates the professional points to.

In this screenshot, the user asked "What do you see here?" by clicking on the inferotemporal quadrant of the retina. The AI responded with a detailed structure:

Soft tissue ultrasound is the first-line study for palpable masses in the neck, extremities, and subcutaneous tissue. When color Doppler is added, the information multiplies: you see not only the shape and size of the lesion, but its vascularity — and that's what separates an innocent lipoma from an aggressive sarcoma.

The following three cases show how RADIA handles Doppler ultrasounds natively, extracting dimensions, vascular pattern, echogenicity, and generating clinical differentials that a musculoskeletal radiologist would validate.

A palpable soft tissue mass. B-mode ultrasound shows a well-defined, predominantly hypoechoic lesion with an oval to lobulated shape. So far, it could be anything. But color Doppler reveals massive hypervascularization — tortuous, abundant vessels with high-energy arterial and venous flow.

The recommendation was clear and highlighted: MRI of the affected area for better delineation + core needle or excisional biopsy for definitive histopathological diagnosis. RADIA doesn't leave the doctor with a "suspicious finding" — it provides the concrete next step.

This case shows two views of the same lesion on color Doppler. The upper image presents an intensely hypervascular mass — a rainbow of red, orange, and blue indicating mixed high-velocity arterial and venous flow. The lower image offers another perspective with a visible scale.

RADIA measured the mass in both views: 35-40 mm × 35 mm in the upper and 15 mm × 10 mm in the lower (visible scale 1.0 unit = 1 cm). It detected irregular contours, tortuous vascularization with a chaotic flow pattern — characteristics "highly suggestive of an aggressive process."

Not everything RADIA detects is malignant. This case shows a complex cystic lesion in subcutaneous tissue — predominantly hypoechoic to anechoic, with multiple thin internal septa and some punctate echoes. No significant internal vascularity on B-mode. Size: 45-50 mm wide by 20-25 mm deep.

The AI classified this finding with Vascular 90% confidence and noted that the margins are somewhat irregular but there are no signs of significant internal vascularity — reducing malignancy suspicion. The screenshot also shows the "Doctor's comment" field, where the professional can add their clinical impression directly on the finding before generating the report.

This case illustrates the importance of AI not over-diagnosing. Not all masses are malignant tumors. A complex cyst requires follow-up but not the urgency of a hypervascular sarcoma. RADIA differentiates between both scenarios with precision, avoiding the cascade of unnecessary tests that a false positive generates.

Behind every finding there's an AI pipeline processing the image in multiple layers. It's not a single model — it's four AI models working in cascade.

RADIA started analyzing dental CTs. Today it processes fundus photographs that detect multicentric retinoblastoma and Doppler ultrasounds that differentiate hemangiomas from sarcomas. The same architecture — four AI models (3 vision + 1 reasoning), Cloudflare Workers, D1, edge analysis — adapts to any clinical image a professional needs analyzed.

What previously required specific desktop software, PACS licenses, and years of subspecialization is now available in a browser. A family doctor uploads a fundus photo taken with a portable retinal camera and gets analysis with bounding boxes, severity levels, clinical differentials, and referral recommendations — in seconds. That's real democratization of medical AI.

Next steps include support for dermatoscopy, obstetric ultrasound, digital mammography, and retinal OCT. The question is no longer whether AI can analyze medical imaging — it's how many specialties can be covered with this same approach.