📄 medicine

Coronal-Plane Feature Analysis in Automated Breast Ultrasound for Differentiating Benign and Malignant BI-RADS 4 Lesions

This study demonstrates that a machine learning model utilizing specific ABUS coronal-plane features, such as retraction phenomenon and ductal changes, achieves expert-level diagnostic performance in differentiating benign from malignant BI-RADS 4 breast lesions, offering a reproducible tool to improve risk stratification and reduce diagnostic variability.

Original authors: Xiajuan Ding, Shangqing Liu, Min Wang, Chengcheng Zhu, Yingxin Lin, Xingping Liu, Jinwei Zhang, Suzhen Ran

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiajuan Ding, Shangqing Liu, Min Wang, Chengcheng Zhu, Yingxin Lin, Xingping Liu, Jinwei Zhang, Suzhen Ran

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Maybe" Problem

Imagine a doctor is looking at a breast lump and has to decide: "Is this a harmless rock (benign) or a dangerous weed (malignant)?"

Usually, the doctor uses a standard ultrasound (like a handheld flashlight) to look. But sometimes, the picture is blurry. The lump looks suspicious enough to worry, but not scary enough to be 100% sure. In the medical world, this confusing middle ground is called BI-RADS 4. It's the "Maybe" zone.

Because the risk varies wildly in this zone (from 2% to 95% chance of cancer), doctors often have to perform a biopsy (a needle test) just to be sure. This study asked: Can a special type of ultrasound called ABUS help us solve this "Maybe" puzzle without needing a needle for every single case?

The Tool: The "3D Movie" vs. The "2D Photo"

Standard ultrasound gives a flat, 2D slice of the breast, like looking at a single slice of bread. It depends heavily on how the doctor holds the wand, which can lead to different opinions.

ABUS (Automated Breast Ultrasound) is different. It's like a robot arm that slowly scans the whole breast and builds a 3D movie.

The Superpower: The real magic of this study wasn't just the 3D view, but a specific angle called the Coronal Plane.
The Analogy: Imagine looking at a tree from the side (standard view). You see the trunk. But if you look at the tree from above (the coronal view), you can see how the roots pull the ground down or how the branches twist. This study focused on looking at breast lumps from this "top-down" angle to see how they interact with the surrounding tissue.

The Detective Work: Five Clues

The researchers taught four radiologists (two experts and two trainees) to look for five specific "clues" in this 3D top-down view:

The Retracting Curtain (Retraction Phenomenon): Imagine a tent pole pulling the fabric inward. If the tissue around the lump is being pulled toward it, it's a strong sign of a "weed" (cancer).
The Broken Pipe (Ductal Changes): Normal milk ducts are smooth tubes. If they look kinked, blocked, or broken near the lump, it's a warning sign.
The Tiny Gravel (Microcalcifications): These are tiny specks of calcium. On this 3D view, they look like bright, sharp pebbles.
The Sharp Corners (Angular Margins): Instead of a smooth, round ball, does the lump have jagged, sharp corners like a star or a claw?
The Glitch (Skipping Sign): This is a weird artifact where the image looks like it "skipped" a beat. The study found this clue wasn't very helpful for this specific group of lumps.

The Results: Who Got It Right?

1. The Experience Gap
The study found that experience matters a lot.

The Trainees: When the two junior doctors looked at the same lumps, they often disagreed. It was like two people trying to describe a cloud; one saw a dog, the other saw a car. Their agreement was low.
The Experts: When the two senior doctors looked, they almost always agreed. They could spot the "pulling curtain" and the "broken pipes" with high consistency.

2. The Clues That Matter
The study confirmed that four of the five clues were powerful predictors of cancer:

If the tissue was being pulled in, the ducts were broken, there was gravel, or the corners were sharp, the lump was very likely to be malignant.
The "Skipping Sign" didn't help much in this specific group.

3. The Computer vs. The Humans
The researchers built a simple computer model (Logistic Regression) that just looked at these five clues to make a decision.

The Trainees: The computer was much better than the junior doctors. It caught more cancers that the trainees missed.
The Experts: The computer performed almost as well as the senior experts. It didn't beat the best expert, but it was right in the same ballpark.

The "Black Box" vs. The "Glass Box"

Usually, computer models are "black boxes"—you put data in, and a result comes out, but you don't know why.

This Study's Model: This was a "glass box." The researchers used a tool called SHAP to peek inside the computer's brain.
The Finding: The computer admitted, "I'm mostly worried about the Broken Pipes and the Pulled Curtain." It confirmed that the computer was using the same logic a human expert would use, making it trustworthy and easy to understand.

The Bottom Line

This study suggests that looking at breast lumps through this special "top-down" 3D view provides clear, repeatable clues about whether a lump is dangerous.

For the "Maybe" (BI-RADS 4) lumps: These specific visual clues are strong indicators of cancer.
For the Doctors: Junior doctors might miss some cancers, but a simple computer model using these clues can help them catch more, acting like a safety net that brings their performance up to the level of experienced experts.

What the study did NOT say:

It did not say this should replace biopsies entirely.
It did not say this works for all types of breast lumps (only the "BI-RADS 4" ones).
It did not claim this is ready for every hospital tomorrow (it was tested in one center and needs more testing).

In short: By looking at the breast from a new angle and using a simple computer helper, doctors can get a clearer picture of those confusing "maybe" lumps, potentially reducing the number of unnecessary needle tests and catching cancers earlier.

Technical Summary: Coronal-Plane Feature Analysis in Automated Breast Ultrasound for BI-RADS 4 Lesions

Problem Statement
Differentiating benign from malignant breast lesions within the BI-RADS 4 category is a significant clinical challenge due to the wide range of malignancy risk (2% to 95%) and substantial interobserver variability among radiologists. While Automated Breast Ultrasound (ABUS) offers standardized, three-dimensional imaging with coronal-plane reconstruction—a modality capable of visualizing spatial relationships between lesions and surrounding tissue—its diagnostic utility specifically for BI-RADS 4 lesions remains under-investigated. Existing literature often evaluates ABUS across heterogeneous BI-RADS categories, lacking specific validation for this indeterminate subgroup. Furthermore, the reproducibility of ABUS coronal-plane features across readers with varying experience levels is not well-established, and there is a need for interpretable tools to assist in risk stratification.

Methodology
This retrospective study analyzed 458 BI-RADS 4 lesions from 373 patients who underwent ABUS prior to biopsy or surgery. The study focused on five specific coronal-plane features:

Retraction phenomenon: Convergence of surrounding tissue toward the mass.
Ductal changes: Abnormalities such as focal dilatation or irregular course.
Microcalcifications: Punctate hyperechoic foci.
Angular margins: Focal, sharp corners forming acute angles.
Skipping sign: Step-like artifacts at the lesion-tissue interface.

Four radiologists (two junior, two senior) independently assessed these features. Interobserver agreement was quantified using Cohen's $\kappa$ . The association between these features and malignancy was evaluated using univariate and multivariate logistic regression.

Four machine learning models—Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Decision Tree (DT)—were developed to classify lesions as benign or malignant based on the five features. Model performance was assessed via stratified five-fold cross-validation. The best-performing model was compared against the diagnostic accuracy of the human readers. Model interpretability was analyzed using Shapley Additive Explanations (SHAP) to quantify feature contributions.

Key Results

Interobserver Agreement: Agreement varied significantly by experience. Among all four readers, agreement was moderate for retraction phenomenon ( $\kappa=0.46$ ) and microcalcifications ( $\kappa=0.47$ ), but poor for angular margins ( $\kappa=-0.03$ ) and low for ductal changes ( $\kappa=0.20$ ). However, agreement among senior radiologists was good to excellent across all features ( $\kappa=0.66–0.96$ ), whereas junior radiologists showed poor to fair agreement ( $\kappa=-0.03–0.28$ ).
Feature Association: In multivariate analysis, retraction phenomenon (OR=24.2), ductal changes (OR=31.2), microcalcifications (OR=15.4), and angular margins (OR=11.2) were independent predictors of malignancy ( $p<0.001$ ). The skipping sign was not a significant predictor.
Machine Learning Performance: The Logistic Regression (LR) model achieved the highest discriminative performance with a mean AUC of 0.87.
- The LR model significantly outperformed junior radiologists (AUC 0.59 and 0.75; $p<0.001$ ).
- The LR model's performance was comparable to Senior Radiologist 2 (AUC 0.85; $p=0.28$ ) but inferior to Senior Radiologist 1 (AUC 0.92; $p<0.01$ ).
- The LR model demonstrated high sensitivity (92%) and Negative Predictive Value (85%), mirroring the profile of experienced readers.
Model Interpretability: SHAP analysis identified ductal changes, microcalcifications, and retraction phenomenon as the primary drivers of malignancy prediction, confirming the biological plausibility of the model's logic.

Significance and Claims
The authors claim that structured assessment of ABUS coronal-plane features provides reproducible and interpretable biomarkers for malignancy risk in BI-RADS 4 lesions. The study demonstrates that a feature-based logistic regression model can achieve diagnostic performance comparable to senior radiologists, offering a potential tool to support risk stratification and enhance diagnostic confidence.

The paper emphasizes that while ABUS coronal-plane features are powerful predictors, their assessment is highly dependent on reader experience, highlighting the need for structured training. The proposed model offers a transparent, explainable alternative to complex "black-box" deep learning approaches, relying on clinically assessable features that can be integrated into human-AI collaboration workflows. The authors conclude that these findings support the use of ABUS-based modeling to reduce the risk of missed or misdiagnosed lesions in this diagnostically challenging category, though they acknowledge the need for future prospective, multicenter validation and automated feature extraction to improve generalizability.