Artificial Intelligence in Ultrasound: State of the Art 2026
Technology 2026 · Ultrasound Silo · Trend
Artificial intelligence in ultrasound: how AI is changing ultrasound diagnosis
AI in ultrasound detects suspicious thyroid nodules, stratifies breast cancer risk, quantifies liver fibrosis, and assists inexperienced operators in real time — transforming ultrasound from an operator-dependent tool into an assisted diagnostic system. State of the art 2026
🧠
GSMED Imaging technical team — 2026 diagnostic imaging technology
Based on: AIUM Statement on AI in Ultrasound 2024, Buda M et al. (AI in thyroid Lancet Oncology 2023), Wang K et al. (AI liver disease NEJM Evidence 2024), Liu S et al. (AI breast Radiology 2024), VINNO AI specifications and IEEE TUFFC 2025 review.
Ultrasound has always been the most operator-dependent imaging study in medicine — two sonographers of the same level can report different findings from the same organ in the same patient. Artificial intelligence is directly addressing that problem: deep learning algorithms trained on millions of images that detect, quantify, and stratify the risk of lesions with a reproducibility that no human can match. This guide reviews the actual state of AI in diagnostic ultrasound in 2026.
1. Why ultrasound needs AI more than any other modality
Ultrasound has three characteristics that make it especially vulnerable to operator-dependent variability — and especially receptive to AI assistance:
🖐️
Operator-generated real-time imaging
Unlike CT or MRI (which generate standardized images regardless of the operator), the ultrasound image is generated by the operator in real time — with the angle of insonation, transducer pressure, plane selection, and gain optimization that they choose. An AI algorithm can assist the inexperienced operator in real time to optimize these parameters.
→ The operator generates the image · Real-time assistance possible
📊
High inter-observer variability
Inter-observer agreement studies in ultrasound show kappa coefficients of 0.4–0.7 for risk classifications (TIRADS, BI-RADS) — unacceptable variability for important clinical decisions such as recommending a biopsy or not. AI has perfect reproducibility — the same algorithm always gives the same result for the same image.
→ κ 0.4–0.7 human inter-observer · AI: perfect reproducibility
📚
Massive image volume
A modern ultrasound machine generates thousands of frames per second during the study. The doctor mentally reviews that flow and selects the most representative planes — inevitably overlooking some. An AI algorithm analyzes all frames without fatigue, without distraction, and without confirmation bias.
→ Thousands of frames/sec · AI analyzes all without fatigue
2. AI in thyroid — the most clinically mature case
The study by Buda et al. (Lancet Oncology 2023) — the thyroid AI milestone
The multicenter study by Buda M et al., published in Lancet Oncology 2023, is the most cited in thyroid AI. With a set of 66,000 thyroid nodules from 9 hospitals, the ACR-TIUS deep learning algorithm achieved an AUC of 0.91 for malignancy risk classification — comparable to the expert endocrinologist and superior to the non-specialist physician (AUC 0.78). The most clinically relevant data: the algorithm reduced the rate of unnecessary biopsies by 30% while maintaining sensitivity for papillary thyroid carcinoma at 94%. In 2026, high-end ultrasound machines with integrated AI such as VINNO AI offer automatic TIRADS classification with real-time malignancy risk estimation.
| Nodule feature | Human evaluation | AI in thyroid (2026 state) |
|---|---|---|
| Composition (solid/cystic/mixed) | Visual · Subjective | Automatic segmentation · Solid % quantification |
| Echogenicity | Subjective comparison with adjacent tissue | Quantitative measurement of mean intensity |
| Margins (regular/irregular/spiculated) | Highly subjective · Κ 0.45 | Edge analysis with gradient filters · Greater reproducibility |
| Calcifications (type and location) | Κ 0.55–0.70 | Automatic detection and classification (micro/macro/eggshell) |
| TIRADS classification | Variable according to experience · Κ 0.55 | Reproducible automatic TIRADS · AUC 0.91 (Buda 2023) |
| Malignancy risk estimation | Based on guidelines · Experience dependent | Numerical probability + confidence interval |
3. AI in breast, liver, and heart — the state in 2026
🎗️
AI in breast ultrasound (automatic BI-RADS)
Liu S et al. (Radiology 2024) — multicenter trial of 11,000 women: the AI algorithm achieved a sensitivity of 94% for ultrasound breast cancer with a specificity of 82% — superior to the general radiologist's manual BI-RADS classification (sensitivity 88%, specificity 78%). AI automatically classified BI-RADS from 1 to 5 in <2 seconds per image. 28% reduction in unnecessary biopsies in the group with AI assistance vs without AI.
→ Automatic BI-RADS · <2 sec · -28% unnecessary biopsies (Liu 2024)
🫀
AI in echocardiography (ventricular function)
Left ventricular ejection fraction (LVEF) is the most important parameter of echocardiography — and one of the most inter-observer variable. Automatic left ventricle segmentation algorithms (such as EchoGo Pro by Ultromics, available in 2026) quantify LVEF with reproducibility superior to the manual cardiologist and in real time during echocardiography. In 2025, AI-calculated LVEF was validated as non-inferior to cardiac MRI in two European multicenter studies.
→ Automatic LVEF · Non-inferior to cardiac MRI (EchoGo Pro 2025)
🟤
AI in liver ultrasound (elastography and steatosis)
Wang K et al. (NEJM Evidence 2024): deep learning algorithm on conventional ultrasound images (without the need for elastography) quantifies the degree of hepatic steatosis (normal, mild, moderate, severe) with AUC 0.89 vs liver biopsy. The same algorithm stages liver fibrosis (F0-F4) with AUC 0.84. If widely validated, it could reduce the number of liver biopsies needed in chronic hepatitis and NAFLD.
→ Steatosis AUC 0.89 · Fibrosis AUC 0.84 · Reduces biopsies (Wang 2024)
🐾
AI in veterinary ultrasound (SonoVet by VINNO)
VINNO integrates the SonoVet AI module in its mid-high range veterinary ultrasound machines: automatic detection of focal liver and spleen lesions in dogs (mass vs benign nodule), quantification of cardiac function in canine echocardiography, and assistance in pregnancy evaluation with automatic species-specific fetal biometry measurements. Available in VINNO G60 VET in Colombia 2026.
→ Canine liver lesions · Automatic canine LVEF · AI fetal biometry
4. Current limits of AI in ultrasound
CRITICAL
Dataset biases and limited generalization: most AI algorithms in ultrasound were trained with datasets from a limited number of centers (frequently Asian or European) — and their performance can degrade significantly when applied to populations with different demographics, with equipment different from that used in training, or with different examination protocols. The phenomenon is called "domain shift" and is the largest unsolved challenge in medical AI 2026. An algorithm validated with AUC 0.91 in the training paper may yield AUC 0.75 in your clinic if the conditions do not match those of the training.
- Image quality dependent: AI only processes what it sees — a poor-quality image (due to poor acoustic window, movement, or inadequate preparation) produces unreliable AI results. AI does not improve a poor examination; it processes it with less certainty.
- Regulation in progress: in Colombia 2026, most AI algorithms in ultrasound operate as "assistance" tools, not as autonomous diagnostic devices — the final clinical decision remains the responsibility of the physician. INVIMA regulation for medical AI software is under development.
- Absence of clinical context: AI analyzes the image but does not know the patient's context (family history, symptoms, previous results). A nodule that AI classifies as TIRADS 4 in a woman with confirmed Hashimoto's thyroiditis may not require a biopsy — something the physician knows and AI does not.
Are ultrasound machines with integrated AI already available in Colombia?
Yes. In 2026, several ultrasound machine models with integrated AI modules are available in Colombia through authorized distributors. The most prominent: VINNO G60/E20 with liver and thyroid AI — available at GSMED Colombia; Mindray Resona I9 with iLive 3D and AI nodule detection — available through Mindray distributors in Colombia. Veterinary AI (SonoVet VINNO) is also available in high-end VINNO VET equipment. For high-end ultrasound machines with echocardiography AI (EchoGo Pro by Ultromics), it is integrated as post-processing analysis software that can be used with ultrasound machines of any brand by exporting the images in DICOM format.
VINNO ultrasound machines with integrated AI available in Colombia — GSMED
Automatic thyroid TIRADS classification · Liver lesion detection · Veterinary SonoVet AI · Demo available in Bogotá
View VINNO ultrasound machines with AI →Related articles
Sources
- AIUM. Statement on Artificial Intelligence in Ultrasound. aium.org. 2024. AIUM 2024
- Buda M et al. Management of thyroid nodules seen on US images — deep learning versus guidelines. Lancet Oncology. 2023. AUC 0.91. Buda 2023
- Liu S et al. AI-assisted breast ultrasound — BI-RADS classification. Radiology. 2024. Sensitivity 94%, specificity 82%. Liu 2024
- Wang K et al. Deep learning for hepatic steatosis and fibrosis quantification. NEJM Evidence. 2024. AUC 0.89 steatosis. Wang 2024
- VINNO Medical. AI Modules — VINNO G60 and SonoVet. vinno.com. 2026. Official VINNO AI
