68/100

Moderate evidence

Cite
More
EmailLinkedInX

An AI model missed no cancers in 90 uncertain breast nodules, a three-hospital study found

Study title: Enhancing diagnostic precision for BI-RADS 4a breast nodules: A multimodal AI model integrating ultrasound radiomics, hemodynamic signatures, and clinical profiles

Evidence type
Controlled or validated comparison
Sample
n = 350
Sources
1 sourcePeer-reviewed
Published
1 min read

Study title: Enhancing diagnostic precision for BI-RADS 4a breast nodules: A multimodal AI model integrating ultrasound radiomics, hemodynamic signatures, and clinical profiles

Researchers developed and validated a multimodal AI model to resolve the diagnostic ambiguity of BI-RADS 4a breast nodules, which have a 2% to 10% malignancy rate and often lead to benign biopsy results. Using data from 350 pathologically confirmed nodules across three hospitals, the LightGBM model combined ultrasound radiomics, hemodynamic features, and clinical profiles, achieving an AUC of 0.95 on a temporally independent test set of 90 nodules as of the October 2026 publication.

Read the study abstract

Background Breast Imaging Reporting and Data System (BI-RADS) 4a nodules represent a diagnostic dilemma, with a malignancy rate ranging from 2% to 10%. The majority of these nodules prove benign after biopsy, leading to unnecessary invasive procedures, patient anxiety, and healthcare costs. Current clinical practice lacks a reliable, noninvasive tool to accurately distinguish benign from malignant BI-RADS 4a lesions. Emerging evidence suggests that integrating multiparametric ultrasound data with clinical factor…

Enhancing diagnostic precision for BI-RADS 4a breast nodules: A multimodal AI model integrating ultrasound radiomics, hemodynamic signatures, and clinical profiles
Hospital Universitari Doctor Peset, València 08 by 19Tarrestnom65. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0

The gain

On 90 held-out breast nodules, a model combining ultrasound, blood-flow and clinical data caught all malignant ones (100% sensitivity) with 86.1% specificity and an AUC of 0.95.

In brief

The result matters because BI-RADS 4a management currently lacks a reliable noninvasive tool, contributing to unnecessary invasive procedures, patient anxiety, and costs. A tool that maintains 100% sensitivity while improving specificity could support clinical decision-making and safely reduce biopsies, though real-world impact, generalizability beyond the three study hospitals, and prospective clinical utility remain to be confirmed.

Main points

  1. Prospective multicenter study enrolled 350 pathologically confirmed BI-RADS 4a nodules from 331 patients in one tertiary hospital and two secondary hospitals.
  2. Model was built with Light Gradient Boosting Machine after LASSO feature selection, using grayscale ultrasound radiomics, hemodynamic parameters, and clinical data.
  3. Training set n=260 with nested cross-validation and hyperparameter tuning; evaluation on temporally independent test set n=90.
  4. Test performance significantly superior to Logistic Regression, Support Vector Machine, and Random Forest baseline classifiers.

The full story

The study collected clinical data, grayscale ultrasound images, and hemodynamic parameters for each nodule and extracted radiomics features, then constructed a multimodal fusion model using LightGBM following LASSO regression for feature selection.

Interpretability was addressed with SHapley Additive exPlanations (SHAP), and the model was developed on 260 nodules and tested on 90 temporally independent nodules, reporting AUC 0.95 with 100% sensitivity and 86.1% specificity.

Sources

1 on file
  1. Peer-reviewed2026-10-01

Talk it over

Three questions to start on, two more if you have time
The model flagged all malignant nodules among the 90 test nodules and had 86.1% specificity. How would you weigh missing no cancers against the nodules it still called suspicious that turned out benign?

Start with a question

An AI model missed no cancers in 90 uncertain breast nodules, a three-hospital study foundhttps://truvace.com/item/enhancing-diagnostic-precision-for-bi-rads-4a-breast-nodules-a-multimodal-ai-mod

  1. The model flagged all malignant nodules among the 90 test nodules and had 86.1% specificity. How would you weigh missing no cancers against the nodules it still called suspicious that turned out benign?

    Answer in the thread
  2. The model was built on 260 nodules and tested on 90 nodules from a later period at the same three hospitals. What can that design show about how it would work somewhere else?

    Answer in the thread
  3. The nodules came from 331 patients at one tertiary hospital and two secondary hospitals. Which patients or clinics might be missing, and how could that change the result?

    Answer in the thread
If you have more time
  • Look up how many malignant nodules were in the 90-nodule test set. How would that number change how much weight you give the 100% sensitivity?

    Answer in the thread
  • If a hospital you know used this model to decide which BI-RADS 4a nodules get a biopsy, who should check its calls and how often?

    Answer in the thread

What readers say