PDF
Citation
Print
INTRODUCTION
Breast cancer remains the most common malignancy among women worldwide, and early, accurate diagnosis is essential for improving outcomes.(1-3) Mammography is the standard screening modality, but its accuracy declines in dense breasts (37-66% prevalence) owing to masking by fibroglandular tissue.(4,5) Breast ultrasound has emerged as an important supplemental screening tool to address these limitations. It demonstrates high sensitivity in women with dense breast tissue, where mammography frequently fails to detect tumors, and it enables real-time imaging that facilitates the detection of small breast cancers.(6-8)
Despite these advantages, conventional breast ultrasound has inherent limitations, as it is highly operator dependent and associated with only moderate inter-observer agreement in Breast Imaging Reporting and Data System (BI-RADS) assessments. Reported diagnostic sensitivity ranges from 67.1% to 87.0% across radiologists.(9,10) Artificial intelligence-based computer-aided detection/diagnosis (AI-CAD) systems has therefore been proposed to reduce variability and enhance accuracy.(11-13) Earlier CAD systems lacked real-time detection and relied on server processing, whereas a real-time solution (CadAI-B® for Breast, version 1.5; BeamWorks Inc., Daegu, Korea) operates on portable devices and generates frame-linked outputs within ~0.04 s at 20-40 fps, enabling on-device clinical decision-making.(11) A recent feasibility study demonstrated that real-time AI-CAD supports lesion detection during breast ultrasonography, contributes to differential diagnosis, and may reduce unnecessary biopsies.(11)
From a clinical perspective, preoperative differentiation between ductal carcinoma in situ (DCIS) and invasive ductal carcinoma (IDC) is critical, as surgical approach and adjuvant therapy often differ substantially between these entities.(14-17) However, DCIS and IDC frequently share overlapping risk factors and imaging appearances, rendering reliable discrimination with conventional ultrasound challenging.(18,19) Ultrasound elastography has demonstrated moderate diagnostic value, particularly in detecting higher shear-wave velocity at lesion margins in IDC compared with DCIS, but overall performance remains limited.(20) In this context, an imaging-based estimate of invasion risk, even when core biopsy yields DCIS, could meaningfully inform surgical planning and multidisciplinary decision-making. Real-time AI-CAD that provides an objective probability of malignancy (POM) and BI-RADS-aligned outputs at the point of care may address this unmet need. Accordingly, we evaluated whether a real-time AI ultrasound system (CadAI-B®) differentiates DCIS from IDC via POM and BI-RADS outputs and which ultrasonographic features drive POM differences.
METHODS
1. Study design and participants
This retrospective, single-center study included consecutive patients who underwent diagnostic breast ultrasonography with concurrent CadAI-B® evaluation at a tertiary medical center. Between January 2023 and December 2023, data obtained from CadAI-B® were retrospectively reviewed. At the time of clinical assessment, all enrolled lesions were assigned ultrasound BI-RADS category 4 by breast radiologists specializing in breast imaging, and 131 such cases were identified. Of these, 97 were confirmed as benign by the reference standard and excluded, and the final study cohort comprised 12 cases of DCIS and 22 cases of IDC. All malignancies were verified by surgical pathology of the resected specimens. For each enrolled case, CadAI-B® was subsequently applied to the ultrasound images to re-evaluate the BI-RADS category and POM. These AI-assigned categories (C1/2, C3, C4A, C4B, C4C, C5) were used for comparative analysis with the final pathological results. Clinicopathologic characteristics, including age, body mass index (BMI), clinical tumor size, presence of suspicious lymph nodes, and type of breast surgery were collected and compared between DCIS and IDC cases. Clinical characteristics for each pathologically confirmed lesion were also recorded, along with the POM generated by CadAI-B®. All data were abstracted from the medical record using a standardized collection form. The study protocol was approved by the Institutional Review Board (IRB) of Kyungpook National University Chilgok Hospital (IRB number: KNUCH 2024-10-022-001).
2. CadAI-B system and ultrasound examination
The CadAI-B® AI platform (version 1.5, BeamWorks Inc., Daegu, Korea) is a tablet-based, real-time diagnostic support tool specifically designed for breast ultrasonography, with prior clinical validation.(11) The platform captures live video feeds from ultrasound devices through HDMI or DVI connections using integrated capture hardware. During scanning, the system continuously analyzes the incoming image stream at a processing speed of 0.04 seconds per frame on CPU architecture, producing visual overlays in the form of heatmaps to highlight regions of potential concern based on computed abnormality indices.
When image acquisition is frozen, the system provides a quantitative malignancy risk assessment expressed as a percentage value (0-100%) and assigns BI-RADS classifications across six categories (C1/2, C3, C4A, C4B, C4C, C5). In this study, POM values were recorded as continuous variables on a 0-1 scale for statistical analysis, corresponding to percentage probabilities (e.g., 0.224 interpreted as 22.4%). This normalization allowed consistent comparison of mean POM values between DCIS and IDC. The platform also performs automated morphometric analysis, determining maximum and minimum lesion dimensions, and generates standardized BI-RADS descriptors encompassing morphology, orientation, boundary characteristics, and echogenic properties (Fig. 1).
The AI architecture employs weakly supervised machine learning to differentiate malignant from non-malignant tissue patterns (including normal parenchyma and benign lesions) while simultaneously identifying suspicious lesions without the need for manual annotation.(21) The neural network’s confidence scoring system was calibrated against expert-annotated reference datasets containing validated BI-RADS categorizations to ensure alignment with malignancy probability thresholds defined in the BI-RADS lexicon.(14)
3. Statistical analysis
Baseline variables were summarized in a descriptive characteristics table. Between-group differences (DCIS vs. IDC) were assessed using independent two-sample t-tests for approximately normally distributed continuous variables, with Welch’s correction applied when variances were unequal. For non-normally distributed variables, the Mann–Whitney U test was used. Categorical variables were compared using the χ² test or Fisher’s exact test when expected cell counts were fewer than five. The primary analysis evaluated POM differences between DCIS and IDC using the same approach as for other continuous outcomes. All statistical tests were two-sided, with P < 0.05 considered statistically significant. Analyses were performed using SPSS Statistics version 21 (IBM Corp., New York, USA).
RESULTS
The mean age was significantly lower among patients with DCIS compared with those with IDC (49.7 ± 6.2 vs. 61.4 ± 11.6 years, P = 0.001), whereas BMI was comparable between groups (22.7 ± 4.0 vs. 23.4 ± 2.7 kg/m², P = 0.633). Tumor size was smaller in DCIS than in IDC (9.8 ± 4.9 vs. 17.0 ± 10.2 mm, P = 0.009). Suspicious lymph nodes were absent in DCIS and present in one IDC case (P = 1.000). There was no significant difference in the type of breast surgery between the two groups (Table 1).
Overall, IDC cases demonstrated a significantly higher mean POM (0.365 ± 0.306) compared with DCIS cases (0.126 ± 0.196, P < 0.001). Several ultrasonographic features were associated with significant POM differences between DCIS and IDC. Almost all lesions were solid masses (97.1%), and IDC demonstrated a significantly higher mean POM than DCIS within this subtype (0.345 ± 0.298 vs 0.126 ± 0.196, P = 0.024). Irregular shape predominated (79.4%) and were associated with a significantly higher POM in IDC than in DCIS (0.344 ± 0.308 vs 0.135 ± 0.203, P = 0.032). IDC lesions with indistinct (0.318 ± 0.226 vs 0.012 ± 0.003, P = 0.017) or microlobulated margins (0.419 ± 0.350 vs 0.132 ± 0.095, P = 0.049) demonstrated significantly higher POM values than DCIS, whereas angular and spiculated margins showed no significant intergroup difference. Most lesions were hypoechoic (94.1%), with IDC showing a higher mean POM than DCIS (0.345 ± 0.298 vs 0.135 ± 0.203, P = 0.023). In terms of orientation to the skin, both groups demonstrated similar distribution patterns (P = 0.476). However, IDC exhibited significantly higher POM in parallel-oriented lesions (0.362 ± 0.290) compared with DCIS (0.044 ± 0.037, P = 0.021). POM differences for non-parallel lesions were not statistically significant (0.367 ± 0.335 vs 0.166 ± 0.233, P = 0.138) (Table 2).
The BI-RADS classification patterns generated by CadAI-B® further highlighted its differential diagnostic capability. IDC cases were predominantly assigned to higher-risk BI-RADS categories, with 40.9% classified as BI-RADS 4B (mean POM: 0.224 ± 0.098), 31.8% as BI-RADS 4C (mean POM: 0.685 ± 0.149), and 4.6% as BI-RADS 5 (POM: 0.963). In contrast, DCIS cases were more frequently assigned to lower-risk categories, with 41.7% classified as BI-RADS 3 (mean POM: 0.021 ± 0.007), 16.7% as BI-RADS 4A (mean POM: 0.088 ± 0.006), and 25.0% as BI-RADS 4B (mean POM: 0.171 ± 0.070) (Table 3).
DISCUSSION
The most notable finding of this study is that IDC demonstrated consistently higher POM values than DCIS across multiple ultrasonographic features. IDC not only exhibited a higher overall mean POM but was also more frequently classified into higher BI-RADS categories, whereas DCIS largely remained within the lower-risk spectrum despite occasional assignment to BI-RADS 4A or 4B. These findings suggest that CadAI-B® may capture visual cues associated with stromal invasion and translate them into quantitative signals reflective of invasive potential.
Several specific differences support this interpretation. Lesions characterized by solid mass, irregular shape, indistinct or microlobulated margins, and hypoechogenicity—all established ultrasonographic markers of malignancy—were consistently associated with higher POM in IDC compared with DCIS. These patterns align with the fundamental biological distinction between non-invasive lesions, which remain confined by the basement membrane, and invasive carcinomas, which disrupt tissue architecture and interact with the surrounding stroma.(22,23) Notably, IDC also demonstrated elevated POM in lesions with parallel orientation, a feature typically considered benign. This observation implies that the model may integrate additional image-level cues, such as subtle contour irregularities or internal heterogeneity, that extend beyond traditional BI-RADS descriptors.
Previous studies have underscored the difficulty of distinguishing DCIS from IDC using conventional ultrasound, given the considerable overlap in imaging features. Elastography and other adjunctive modalities have been investigated, though their diagnostic performance remains modest.(20,24,25) In this context, the present findings suggest that AI-derived POM may provide complementary diagnostic value by systematically capturing ultrasound features associated with invasive disease. While BI-RADS classification remains the standard reference for biopsy decision-making, incorporating AI-derived probabilities has potential to enhance preoperative counseling, surgical triage, and multidisciplinary planning—particularly in cases where a biopsy diagnosis of DCIS raises concern for possible underestimation of invasive disease.
The real-time processing capability of CadAI-B® also warrants emphasis. Earlier AI-based ultrasound solutions were constrained by non-real-time inference or server dependence.(11,26) In contrast, CadAI-B® generates frame-linked outputs within approximately 0.04 seconds per frame, a speed appropriate for examinations that produce tens of thousands of frames. This immediate feedback minimizes the risk of overlooking critical visual cues and may help reduce the variability known to occur among human interpreters. By providing standardized, objective findings during scanning, the system may enhance both lesion detection and risk stratification in routine practice.
These findings should be interpreted in light of study limitations. The retrospective design and relatively small sample size limit generalizability. Another important limitation involves potential false negatives in AI classification. Although all enrolled lesions were clinically assessed as BI-RADS 4 at ultrasound, several cases were reclassified by CadAI-B® into lower AI categories (C1/2 or C3) at the freeze frame. In our center, a subset of these C1/2/3 cases underwent surgery following core-needle biopsy confirmation; such cases may represent “AI false negatives.” To minimize this issue, future studies should analyze multiple representative frames per lesion rather than relying on a single freeze-frame, perform threshold-sensitivity and calibration analyses to refine decision thresholds, establish collaborative reader–AI decision protocols that recommend biopsy when the clinician’s BI-RADS is ≥4 regardless of the AI-derived POM, and conduct prospective validation using standardized biopsy and surgical criteria. Nonetheless, several strengths are notable. All cases were histopathologically confirmed and supported by comprehensive clinical and pathological documentation, enabling reliable comparative analysis between DCIS and IDC. Moreover, the evaluation included not only overall POM but also feature-level differences, offering a nuanced perspective on how AI systems interpret classical ultrasound descriptors.
In summary, the findings indicate that CadAI-B® has potential to complement conventional ultrasound by providing quantitative outputs that correspond with invasive tumor biology. Although larger, prospective studies are needed to validate these preliminary results and to establish clinical impact, the present data suggest that real-time AI support may play a meaningful role in preoperative risk assessment and decision-making.
CONCLUSION
The real-time AI ultrasound system CadAI-B® demonstrates potential utility in distinguishing DCIS from IDC during breast ultrasonography. By leveraging ultrasonographic features, the system provides distinct POM outputs and BI-RADS classifications, with IDC consistently exhibiting higher POM values and being assigned to higher BI-RADS categories compared with DCIS. This capability may serve as a valuable tool for improving clinical classification, facilitating immediate decision-making, and delivering consistent, objective data to optimize patient care. Although these preliminary findings are encouraging, large-scale multicenter clinical trials are necessary to validate the diagnostic value of CadAI-B® in broader clinical settings and to fully realize its potential in enhancing breast cancer diagnosis.
ACKNOWLEDGMENTS
This study was supported by a grant (IRB number: KNUCH 2024-10-022-001) from Kyungpook National University Chilgok Hospital.
ETHICAL APPROVAL STATEMENT
All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.
XML Download