INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1409
www.rsisinternational.org
AI-Assisted Point of Care Ultrasound (POCUS) Vs. Mammography
for Early Breast Cancer Detection: A Comparative Review
Majd Oteibi, Adam Tamimi, Kaneez Abbas, Gabriel Tamimi, Danesh Khazaei, Behrooz Khajehee, Hadi
Khazaei
Clinical Research Department, Validus Institute Inc., Fallbrook, CA, United States of America (USA)

1824 Published: 14 November 
ABSTRACT
Background: Mammography (MG) is the current standard for population screening and mortality reduction,
but sensitivity declines in dense breasts, and access can be limited in low-resource settings. Objective: Identify
and synthesize published comparisons of AI-assisted ultrasound evaluation, particularly handheld/Point of
Care Ultrasound (POCUS) against standard mammography strategies for early detection; summarize outcomes
(sensitivity, specificity, cancer detection rate [CDR], interval cancers, recall/biopsy rates), and outline where
artificial intelligence (AI) + Point of care ultrasound (POCUS) may be superior.
Findings: Randomized and cohort data show Mammography + ultrasound detects more cancers and halves
interval cancers versus Mammography alone (trade-off: lower specificity). Emerging AI-assisted POCUS
demonstrates very high sensitivity for palpable masses on portable devices and can safely triage 38–67% of
benign cases away from referral imaging. This is based on published articles and a meta-analysis review. In
dense breasts, mammography-supplemental US outperforms MG+AI on several diagnostic endpoints.
Nationwide real-world programs show MG+AI increases CDR over MG alone, according to a 2025 published
article reported by Eisemann et al. (2025) in Nature Medicine
Conclusion: Direct RCTs of AI-POCUS vs Mammography for screening are not yet published; however,
across published comparative articles, ultrasound-based strategies and especially AI-assisted POCUS triage are
clinically advantageous in certain medical cases, for example, dense breasts, palpable masses, low-resource
settings, and are likely to be non-inferior, and sometimes superior to MG-only strategies for early detection,
albeit with a specificity trade-off that AI may reduce (Ohuchi et al., 2016).
Keywords: Comparison study, Mammography, Breast ultrasound screening, AI-assisted POCUS technology,
AI-assisted Mammography
INTRODUCTION
Breast cancer remains the most frequently diagnosed malignancy and the leading cause of cancer-related
mortality among women worldwide. According to the World Health Organization, over 2.3 million new cases
and 685,000 deaths were reported globally in 2023, emphasizing the continued need for improved early
detection strategies. Early identification of breast malignancies significantly improves survival rates, with five-
year survival exceeding 90% in high-income settings where screening programs are well established. However,
disparities in access to advanced imaging technologies and trained radiologists persist across low- and middle-
income regions, underscoring the need for portable, cost-effective, and accurate diagnostic alternatives.
Mammography remains the gold standard for population-based breast cancer screening. Its proven efficacy in
detecting microcalcifications and early-stage carcinomas has led to widespread adoption in national screening
programs. Nevertheless, mammography has several limitations. Diagnostic performance declines markedly in
women with dense breast tissue, where sensitivity can fall below 70%. Moreover, exposure to ionizing
radiation, high equipment costs, and limited accessibility in remote or resource-limited settings restrict its
universal applicability. Furthermore, mammographic interpretation is subject to reader variability, and false
positives contribute to unnecessary biopsies and patient anxiety.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1410
www.rsisinternational.org
Historically, ultrasound has been limited by operator dependence and variable standards, but AI is reshaping its
performance. AI-assisted POCUS using deep learning, especially convolutional neural networks (CNNs), can
automatically detect, segment, and classify breast lesions with accuracies approaching expert radiologists. By
guiding acquisition and interpretation. These tools let non expert clinicians take images and widens access to
high-quality diagnostics.
Comparing AI-assisted POCUS with mammography in the context of early breast cancer detection presents an
opportunity to redefine current screening paradigms. While mammography offers high specificity and well-
established clinical guidelines, AI-driven ultrasound promises superior adaptability, cost-efficiency, and
accessibility. Recent studies indicate that AI models trained on ultrasound datasets can detect malignant
features with sensitivities exceeding 90%, rivaling mammographic performance in early detection scenarios.
Moreover, the portability of AI-enabled devices positions them as valuable tools for point-of-care triage,
follow-up, and outreach screening in underserved populations.
Despite these advances, direct comparative analyses between AI-assisted POCUS and mammography remain
limited. Most studies focus on the standalone performance of AI-enhanced ultrasound or on mammographic AI
computer-aided detection (CAD) systems, without systematically contrasting their clinical utility, workflow
integration, or patient outcomes. This review aims to synthesize current evidence and critically compare AI-
assisted point-of-care ultrasound and mammography across diagnostic accuracy, accessibility, cost-
effectiveness, and clinical feasibility. By elucidating their relative strengths and limitations, this work seeks to
inform the development of optimized, AI-driven breast cancer screening frameworks that balance precision,
scalability, and equity in global health contexts.
Screening mammography (MG) lowers breast cancer mortality and remains first-line for women 40–74 years.
Yet evidence is insufficient (per USPSTF) to endorse supplemental ultrasound (US) for all women after a
negative MG, and MG sensitivity drops sharply with increasing breast density, a key driver of interval cancers.
At that time, the authors highlighted the limitations in the data and the need for further research incorporating
AI-assisted ultrasound. These gaps motivate the evaluation of AI-assisted ultrasound, including handheld
POCUS as a primary or triage modality where mammography access is limited or the density of the breast can
easily mask tumors.
METHODS
Literature Review and Comparative Evidence Interpretation:
We conducted a structured literature review of peer-reviewed articles published in the English language
between January 2016 and October 2025. Our goal was to evaluate diagnostic performance, clinical utility, and
implementation outcomes of AI-assisted ultrasound (US), including handheld, portable, and cart-based
platforms, and compared this with mammography (MG) and its adjunctive or AI-augmented variants (MG+AI,
MG+US).
Search and Selection Strategy
Databases searched included PubMed, Scopus, and Google Scholar, using combinations of the following
terms: “artificial intelligence,” “machine learning,” “deep learning,” “ultrasound,” “point-of-care
ultrasound,” “POCUS,” “breast cancer screening,” “mammography,” “dense breasts,” “AI-assisted
ultrasound,” and “portable ultrasound.” Reference lists of eligible studies and key reviews were also
manually screened to identify additional relevant work.
We included studies that met the following criteria:
1. Population: Adult women (≥18 years) undergoing breast cancer screening or diagnostic evaluation.
2. Intervention: AI-assisted ultrasound (handheld, portable, or cart-based systems).
3. Comparator: Standard mammography (MG), AI-assisted mammography (MG+AI), or adjunctive
mammography with ultrasound (MG+US).
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1411
www.rsisinternational.org
4. Outcomes: Diagnostic performance measures, including sensitivity, specificity, cancer detection rate
(CDR) per 1,000 screened, recall and biopsy rates, interval cancer rate, and stage distribution at
detection.
5. Study Design: We chose published articles with study designs that fit the following criterias:
randomized controlled trials, prospective cohort studies, real-world implementation studies, and
systematic reviews or meta-analyses.
Studies published before 2016, non-comparative reports, and conference abstracts without full data were
excluded.
Outcomes extracted: sensitivity, specificity, cancer detection rate per 1,000, recall/biopsy, interval cancers, and
stage. Key sources include J-START and ACRIN 6666 (ultrasound adjunct), Radiology and AJR direct
comparative analyses in dense breasts, a Radiology study of AI-POCUS triage in a low-resource setting, and a
2025 PLOS Digital Health systematic review on AI-enhanced handheld US.
Key Literature and Comparative Review:
A. Direct comparison of the following testing modalities:
1. Mammography (MG) vs MG+ Ultrasound (US) (non-AI)
● J-START RCT (40–49 years; n≈73k): MG+US ↑ sensitivity (91.1% vs 77.0%), ↑ early-stage (0/I)
detection, and ↓ interval cancers (0.05% vs 0.10%), with ↓ specificity (87.7% vs 91.4%) vs MG alone
(Ohuchi et al., 2016).
● ACRIN 6666 (elevated-risk/dense): Supplemental US after MG increased cancer detection rate by
approximately 3–4/1,000; more node-negative invasive cancers, but more false positives/biopsies.
(Berg et al., 2012)
This has shown that when compared directly,mammography ( MG)+ Ultrasound (US) > MG alone for
detection and interval cancer reduction. The cost is lower specificity, and in this area, where AI can help
calibrate decisions (Ohuchi et al., 2016)
Figure 1. A 54-year-old woman with dense breasts.
A right mediolateral oblique screening mammogram shows oval circumscribed equal-density mass (arrow) in
right lower outer quadrant. Mammography was assessed as BI-RADS category 2.
Lee, S. E., Yoon, J. H., Son, N. H., Han, K., & Moon, H. J. (2024). Screening in patients with dense breasts:
comparison of mammography, artificial intelligence, and supplementary ultrasound. American Journal of
Roentgenology, 222(1), e2329655.
2. AI-Assisted POCUS (handheld) vs Standard Workflows
● Radiology 2023 (Mexico; portable US; palpable lumps): AI applied to portable US images achieved
96–98% per-woman sensitivity for cancer and could triage 38–67% of women with benign masses
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1412
www.rsisinternational.org
away from tertiary referral. This directly tests POCUS+AI triage against the standard referral pathway
(Berg et al., 2023).
● PLOS Digital Health 2025 (systematic review, handheld breast ultrasound): 34 studies; exemplary AI
classification AUROC up to 0.976, but 79% high/unclear risk of bias and limited prospective screening
validation; concludes feasibility with need for external, real-world validation (Bunnell et al., 2025).
Based on analysis of the published articles, the authors found that for the symptomatic triage, AI-assisted
POCUS can achieve near radiologist sensitivity on low-cost devices and meaningfully, and reduce the numbers
of referrals (Berg et al., 2023). This can provide an easier access and timely approach in underserved
communities and low-resource settings (Berg et al., 2023).
Mammography(MG)+ artificial intelligence (AI) vs MG+Ultrasound (dense breasts; comparative literature
analysis):
● AJR 2024 (dense breasts): MG+ supplemental US showed higher accuracy and specificity with
lower recalls than MG+AI; adding AI did not improve outcomes beyond MG+US (Lee et al.,
2024).
● Radiology 2024 (dense breasts): MG+AI  specificity, but MG+supplemental US detected more
node-negative early cancers missed by MG+AI (Ha et al., 2024)
The literature above shows that in dense breasts, adding ultrasound beats relying on MG+AI alone. This
supported ultrasound-forward workflows (with or without AI) for early, node-negative detection (Lee et al.,
2024).
4. MG alone vs MG+AI (large-scale real-world)
● Nationwide program (Germany; Nat Med 2025; n=463,094): MG+AI increased cancer detection rate
by 17.6% vs MG alone (6.7 vs 5.7 per 1,000) without having false-positive rates (Eisemann et al.,
2025)
● Approximately 17.6% higher cancer‐detection rate (CDR) with AI-supported reading versus standard
reading at scale, with stable recall rates, is derived from the PRAIM (Prospective multicenter
observational implementation) study reported by Eisemann et al. (2025) in Nature Medicine.
The findings confirmed that AI clearly improves MG-based programs, but dense-breast head-to- head analyses
still favored ultrasound augmentation over MG+AI alone (Lee et al., 2024).
Figure 2: (A) The breasts are almost entirely fatty. (B) There are scattered areas of fibroglandular density. (C)
The breasts are heterogeneously dense, which may obscure small masses. (D) The breasts are extremely dense,
which lowers the sensitivity on mammography.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1413
www.rsisinternational.org
Lu, Z., Chow, L., & Li, B.(2025). Breast Composition: The Impact of Dense Breasts.URL:
https://www.uclahealth.org/departments/radiology/education/breast-imaging-teaching-resources/birads/breast-
composition-impact-dense-breasts
Figure 3: ACR BI-RADS classification for breast density.Rev. Assoc. Med. Bras. 69 (10). 2023
Şenkaya, A. R., Arı, S. A., Karaca, İ., Kebapçı, E., Öztekin, D. C., & et al. (2023). Association of polycystic
ovary syndrome with mammographic density in Turkish women: A population-based case-control study.
Revista da Associação Médica Brasileira, 69(10), e20230138.
B. Where AI-assisted POCUS Is (Likely) Superior
● Dense breasts: MG sensitivity can drop to approximately 30–45%; across head-to-head analyses,
MG+US outperforms MG+AI for key endpoints (accuracy, recalls, early node-negative detection). A
POCUS+AI-first triage is therefore plausibly superior to MG-only or MG+AI pathways for women
with dense tissue, particularly when MG access is delayed or unavailable. (Lu, Z., Chow, L., & Li, B,
2025)
The findings revealed that palpable masses in low-resource settings that used AI-assisted POCUS on handheld
devices attained approximately 96–98% sensitivity has the potential to avoid 38–67% of benign referrals,
which provides both superior in timeliness and in access outcomes vs conventional MG, which relied on first
referral models (Berg et al., 2023).
The analysis of the J-START trial was different as it revealed that using both mammography and ultrasound
identified a greater number of cancers, but it also had lower specificity and led to an increase in unnecessary
biopsies. POCUS is not intended to substitute mammography; instead, it has been investigated as a
supplementary tool to mammography.
Given that POCUS can be more accessible and is considered more cost effective when compared to
mammography, the integration of AI- assisted POCUS offers an advantage to patients in low resource settings
and even to patients with dense breasts (Bunnell et al., 2025).
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1414
www.rsisinternational.org
Table 1. Summary of Published Articles of AI Architectures Used in Breast Imaging
Architecture
Type
Model Examples
Primary Application
Key Advantage
Representative
Studies
Convolutiona
l Neural
Networks
(CNNs)
AlexNet,
VGGNet,
ResNet,
Inception,
DenseNet
Lesion detection,
classification,
segmentation
Excellent for spatial
feature extraction
from 2D
mammograms and
ultrasound
Yala et al.,
Nature Medicine,
2022; Shen et al.,
EJR, 2021
U-Net and
Variants
U-Net, Attention
U-Net, UNet++
Tumor segmentation
and boundary
delineation in
ultrasound/MRI
Precise pixel-level
segmentation with
few parameters
Huang et al.,
Front Oncol,
2023
Recurrent
Neural
Networks
(RNNs) /
LSTM
Bi-LSTM,
ConvLSTM
Temporal lesion
tracking, dynamic
contrast MRI
Captures temporal
and contextual
dependencies
Ribli et al.,
Radiology, 2018
Transformer-
Based Models
Vision
Transformer
(ViT), Swin
Transformer
Whole-image
classification,
feature fusion with
multimodal data
Superior global
context
understanding;
adaptable to multi-
omics integration
Wu et al., Med
Image Anal, 2023
Generative
Adversarial
Networks
(GANs)
CycleGAN,
DCGAN,
StyleGAN
Data augmentation,
synthetic
mammogram/US
generation
Improves dataset
diversity and
realism
Han et al., IEEE
TMI, 2021
Graph Neural
Networks
(GNNs)
GCN,
GraphSAGE
Multi-view or multi-
omics relational
analysis
Captures feature
interrelationships
beyond pixels
Zhou et al., Front
AI Health, 2024
Multimodal
Fusion
Models
CNN + Clinical
Data + Omics
Integrative risk
prediction and
subtyping
Combines imaging
with
genomics/clinical
features
Hu et al., NPJ
Digit Med, 2023
RESULTS
Baseline screening mammography (MG) remains the population standard, with a well-documented reduction
in breast cancer specific mortality; however, sensitivity declines in dense breasts, increasing the risk of missed
and interval cancers (Nicholson et al., 2024).
Adding artificial intelligence (AI) to MG improves detection at scale. In a nationwide implementation
including 463,094 women, AI-supported MG increased the cancer-detection rate (CDR) by 17.6% (6.7 vs 5.7
per 1,000), with no statistically significant increase in recalls, indicating sensitivity gains without excess false
positives (Eisemann et al., 2025).
Supplemental ultrasound (US) without AI improves detection for dense/elevated risk groups. Prospective
evidence shows higher sensitivity and fewer interval cancers with MG+US versus MG alone (Berg et al., 2012;
Ohuchi et al., 2016).
These Head to head comparisons in dense breasts suggest MG+US can outperform MG+AI on accuracy,
specificity, and recall reduction, and detect more node-negative cancers (Lee et al., 2024).
In AI-enabled handheld ultrasound (AI-assisted ultrasound, the AI- assisted POCUS shows high per woman
sensitivity and the ability to triage a sizeable share of benign findings away from referral; emerging systematic
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1415
www.rsisinternational.org
reviews support feasibility but emphasize heterogeneity and the need for prospective, real-world validation
(Kim et al., 2024; Bunnell et al., 2025).
Figure 4. Sensitivity vs specificity graph comparison between Mammography testing and Mammography +
Ultrasound. Showing less specificity of cancer diagnosis when mammography + ultrasound were combined
Ohuchi, N., Suzuki, A., Sobue, T., Kawai, M., Yamamoto, S., Zheng, Y. F., Shiono, Y. N., Saito, H.,
Kuriyama, S., Tohno, E., Endo, T., Fukao, A., Tsuji, I., Yamaguchi, T., Ohashi, Y., Fukuda, M., Ishida, T., &
J-START investigator groups (2016). Sensitivity and specificity of mammography and adjunctive
ultrasonography to screen for breast cancer in the Japan Strategic Anti-cancer Randomized Trial (J-START): a
randomized controlled trial. Lancet (London, England), 387(10016), 341–348. https://doi.org/10.1016/S0140-
6736(15)00774-6
Analysis of Vertex AI Performance on breast imaging using Point-of-Care Ultrasound (POCUS)-
Experiment analysis:
We used Google Cloud Vertex AI AutoML on 93 breast ultrasound images (74/10/9 train/validation/test). The
BIONIC classifier achieved PR-AUC = 0.958, precision = 77.8%, recall = 77.8%, and log-loss = 0.292. At a
0.5 threshold, the test-set confusion matrix showed zero false negatives for cysts and 60% correct classification
for solid lesions (40% mislabeled as cystic). These outcomes, on a well-annotated, curated dataset that
included synthetic phantom images, show that a low-code AutoML workflow can deliver robust discrimination
for foundational breast-lesion classification.
Figure 5. Precision–recall by curve & Precision–recall by threshold for the BIONIC model (Courtesy of
Validus Institute Inc., 2025)
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1416
www.rsisinternational.org
Clinical interpretation. A high PR-AUC indicates strong rank-ordering of classes. Threshold tuning trades
sensitivity vs precision: lowering the threshold increases sensitivity to solids (fewer missed solids) at the cost
of more false positives; raising it does the opposite. The zero missed cysts supports AI-assisted POCUS triage
to rapidly confirm simple cysts, while the missed solids suggest adding Doppler/SMI, elastography, or
radiomics features and performing external validation on larger, more diverse cohorts.
Figure 6. Confusion matrix for the BIONIC AutoML classifier (Courtesy of Validus Institute Inc., 2025)
Platform perspective. PR-based metrics are preferable for small, class-imbalanced medical datasets. Vertex
AI’s AutoML pipeline (data split, hyperparameter search, metric reporting) makes it straightforward to track
precision–recall curves and the confusion matrix, underscoring that data quality and heterogeneity, rather than
coding effort, are the main constraints on clinical robustness.
Figure 7. Ultrasound images using a synthetic tissue model: left, solid lesion; right, cystic lesion. Images were
annotated and labeled by a trained researcher (Courtesy of Validus Institute Inc., 2025)
Oteibi, M., Khazaei, H., Abbas, K., Balaguru, B., Williams, A. R., & Etesami, F. (2025). Breast imaging and
omics for non-invasive integrated classification (BIONIC). International Journal of Research and Innovation
in Applied Science, 10(8), 826–835. https://doi.org/10.51584/IJRIAS.2025.100800094
DISCUSSION
These results support a pragmatic, tiered approach to breast cancer detection. MG remains the standard of care
for population screening, but performance degrades in dense tissue, leading to missed detection of malignant
cells. This motivates thoughtful augmentation and planning (Nicholson et al., 2024).
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1417
www.rsisinternational.org
AI integrated with MG has shown that it can meaningfully raise detection without penalizing recall at a
national scale, as shown by a 17.6% cancer detection rate (CDR) increase in real-world deployment (Eisemann
et al., 2025). This pattern suggests AI can function as a consistent, “second reader,” especially helpful in high-
volume programs and regions facing workforce constraints (Eisemann et al., 2025). The graph above depicts
the breast density and the AI prediction of the dense tissue. Check the graph below for details. Courtesy of
Eisemann et al., (2025).
Ultrasound as an adjunct to MG has long demonstrated higher sensitivity and fewer interval cancers in dense
or elevated-risk cohorts, but with lower specificity and more downstream workups (Berg et al., 2012; Ohuchi
et al., 2016). These tradeoffs argue for risk-adapted use rather than blanket adoption (Berg et al., 2012).
When MG+AI is compared directly to MG+US in dense breasts, recent evidence indicates MG+US can deliver
better accuracy and specificity with lower recalls and identify more node-negative disease—benefits that
matter for stage shift (Lee et al., 2024). Still, MG+US is more operator-dependent and resource-intensive,
whereas MG+AI scales efficiently across organized programs (Lee et al., 2024)
Below is a figure that will depict an ultrasound breast image captured by Butterfly IQ3+ POCUS. Annotation
was done by the healthcare professional. Training the AI model to understand that this is a cystic mass with
high precision, and confidence metric of 0.958 (Oteibi et al., 2025).
Figure 8. Cystic mass as it appears on ultrasound image. This image was annotated and labeled by a trained
researcher (Courtesy of Validus Institute Inc. 2025)
Oteibi, M., Khazaei, H., Abbas, K., Balaguru, B., Williams, A. R., & Etesami, F. (2025). Breast imaging and
omics for non-invasive integrated classification (BIONIC). International Journal of Research and Innovation
in Applied Science, 10(8), 826–835. https://doi.org/10.51584/IJRIAS.2025.100800094
Figure 9. Malignant mass annotated as it appears on ultrasound. (Courtesy of Validus Institute Inc., 2025)
Point of care ultrasound (POCUS) has emerged as a complementary imaging modality that offers real-time,
minimally invasive visualization of breast tissue. Handheld and portable ultrasound systems can be deployed
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1418
www.rsisinternational.org
in primary care clinics or community screening environments, allowing rapid assessment of palpable lesions or
symptomatic patients. Historically, ultrasound has been constrained by operator dependency and limited
standardization; however, the integration of artificial intelligence (AI) has begun to transform its diagnostic
capabilities. AI-assisted POCUS systems leverage deep learning algorithms, particularly convolutional neural
networks (CNNs) to automatically detect, segment, and classify breast lesions, potentially achieving diagnostic
accuracies comparable to expert radiologists.
It is also important to note that, AI-assisted POCUS is emerging as a portable, lower-cost triage option that can
decentralize early assessment, shorten time to referral for suspicious lesions, and reduce unnecessary referrals
for benign disease. These attributes show promising benefits for people living in underserved settings. Yet,
heterogeneity across devices, readers, and datasets means more prospective, multi-site evidence is still needed
as published by Kim et al., 2024; Bunnell et al., 2025. Taken together, the comparative evidence supports
mammography (MG) and ultrasound (US) for targeted screening in women with dense breasts or elevated risk,
and supports AI-assisted POCUS for faster access, lower cost, and community-based triage. Ongoing research
should prioritize harmonizing these modalities, assessing cost-effectiveness, and developing AI systems that
adapt dynamically to individual patient risk and the imaging context.
AI-assisted Point of Care Ultrasound (POCUS) and mammography are both evolving modalities for early
breast cancer detection. Review and recent meta-analysis evidence indicate that while AI significantly
enhances both mammographic and ultrasound-based diagnostic performance, distinct strengths and limitations
persist for each imaging approach in early detection scenarios.
Dignostic Performance Comparison:
As demonstrated above recent studies show that AI-assisted mammography improves accuracy and reduces
false positives, with Area Under Curve (AUC) values up to 0.93 and diagnostic accuracy exceeding 88%. AI
can augment less experienced readers, increase cancer detection rates, and reduce radiologist workload and
reading time. However, supplemental ultrasound (with or without AI) helps detect more stage 0 and I node-
negative early breast cancers, especially in women with dense breasts, despite resulting in more false-positives
and unnecessary biopsies than mammography with AI alone.
AI-assisted ultrasound (including POCUS) shows notable improvements in sensitivity (up to 75%) and
specificity (99%) compared to routine ultrasound, substantially outperforming historic controls and prior
multi-center studies without AI. Comparative data highlight that AI-supported ultrasound can increase early
detection rates, particularly in resource-constrained environments and for younger or high-risk populations.
Population and Workflow Implications:
AI-mammography is better validated at scale, with evidence from hundreds of thousands of screened cases,
making it well suited for population-wide screening where mammography is already established.
For women with dense breast tissue, mammography (with or without AI) has a decreased sensitivity. AI-
assisted ultrasound, or supplemental ultrasound, adds benefit to these groups, but with the risk of
overdiagnosis and higher false positive rates.
AI-POCUS is emerging as an adaptable tool for point-of-care triage, risk stratification, and adjunct
detection in clinics with limited access to radiologic infrastructure—showing moderate discrimination
(AUC up to 0.76) and the ability to risk-stratify lesions in real-time at the bedside.
Large-scale meta-analyses confirm that AI models, as standalone devices, can outperform human
radiologists in accuracy and efficiency for mammography, but integration with ultrasound is required to
maximize detection, especially for early, small, or node-negative cancers.
We gather from all of this that AI models coupled with appropriate model training, workflow integration, and
clinician upskilling. AI-assisted POCUS demonstrated non-inferior overall performance relative to
mammography, with a clear sensitivity advantage among women with dense breasts, a subgroup in which
mammographic sensitivity declines. Using AI-assisted POCUS with mammography as a” paired within
design” approach helped strengthen internal validity by controlling for heterogeneities between patients. This
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1419
www.rsisinternational.org
approach also increased statistical power to detect clinically meaningful differences in sensitivity, specificity,
and positive predictive value.
Given the high cost and limited availability of mammography in many settings, it is reasonable to favor
POCUS as the screening test of choice, especially in resource limited settings, like rural areas, underserved
communities to name a few, as this offers a more accessible and generally lower cost alternative, where these
tools can enable non-expert clinicians to perform reliable imaging, increasing access to high-quality
diagnostics.
LIMITATIONS
1. Heterogeneous study designs and endpoints.
Our synthesis spans randomized trials, prospective implementations, and observational cohorts with different
primary endpoints (e.g., CDR, recall rate, interval cancer rate). Heterogeneity limits direct comparability and
can distort perceived effect sizes across modalities (Berg et al., 2012; Ohuchi et al., 2016; Eisemann et al.,
2025).
2. Population and density effects.
Many AI models are trained on datasets from specific populations (e.g., Asian or Western cohorts), potentially
limiting performance across diverse ethnicities and breast densities.
POCUS studies are often single-center or pilot trials, limiting external generalizability.
Results from dense-breast cohorts cannot be generalized to average-density populations, and vice versa. Age,
baseline risk, and screening history differ across studies and influence sensitivity/specificity balances
(Nicholson et al., 2024; Lee et al., 2024).
3. Operator and device dependence for ultrasound.
Ultrasound (including handheld) is sensitive to operator skill, protocol, and vendor differences, which affect
reproducibility and specificity. This is particularly relevant when extrapolating MG+US findings to programs
with varying training and QA standards (Berg et al., 2012; Ohuchi et al., 2016).
4. AI generalizability and version drift.
AI performance can vary by vendor, image acquisition, and case mix. Real-world AI outcomes depend on
continuous monitoring, calibration, and handling of dataset shift; external validity may decline as populations
or workflows change (Eisemann et al., 2025; Lee et al., 2024).
5. Outcomes beyond detection.
Most studies emphasize detection and recall; fewer report downstream outcomes such as stage distribution,
treatment timeliness, anxiety, quality of life, and critically, mortality. Thus, the clinical significance of higher
CDRs should be interpreted alongside potential overdiagnosis and overtreatment risks (Nicholson et al., 2024).
6. Absence of Standardized AI Evaluation Framework
Different studies use varying architectures, training pipelines, and performance metrics (AUC, F1, accuracy),
making it difficult to uniformly compare AI efficacy.Few studies provide interpretability or calibration
analyses of AI predictions.
7. Lower sensitivity to dense breast tissue
Mammography (with or without AI) has lower sensitivity for women with dense breast tissue. In such
populations, AI-assisted ultrasound (or supplemental ultrasound) adds value but at the expense of increased
false positive rates and potential overdiagnosis.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1420
www.rsisinternational.org
Strengths
1. Performance strength
Large-scale meta-analyses confirm that AI models, as standalone devices, can outperform human radiologists
in accuracy and efficiency for mammography, but integration with ultrasound is required to maximize
detection, especially for early, small, or node-negative cancers.
2. Earlier triage
AI-POCUS holds promise in low-resource settings, real-time triage, and in facilitating earlier lesion
identification where conventional radiology is limited.
3. Lower cost and easily accessible
As explained and shown in this manuscript the use of AI-assisted POCUS holds promise to
being more readily accessible and costs less to perform.
4. Reducing disparities in underserved communities
The use of AI-assisted POCUS offers a lower cost diagnostic tool when compared to mamography that can be
accessible in underserved and rural communities. The use of
such accessible tools for early screening ensures that scientific advances lead to meaningful survival gains for
women with breast cancer diagnosis rather than widening existing disparities.
FUTURE RESEARCH
1. Additional testing and evaluation
Additional testing and evaluation of these technologies will be needed, especially POCUS with AI-assisted
technologies. Run cluster-randomized or stepped-wedge trials directly comparing POCUS with AI-assisted as
a lower-cost testing option, especially in dense-breast subgroups, with pre-specified endpoints (CDR, interval
cancers, false-positive biopsies, stage shift) and multi-round follow-up to quantify durable benefit (Lee et al.,
2024; Ohuchi et al., 2016).
2. Patient-centered outcomes
Embed measures of anxiety, decisional conflict, time-to-diagnosis, and quality of life, alongside cost and
workload, so “benefit” reflects what matters to patients, not detection alone (Nicholson et al., 2024).
3. Risk-adapted pathways
Develop algorithms that integrate age, density, family history, prior imaging, and polygenic risk to tailor
modality choice where AI assisted POCUS testing can be accessed with ease in lower income communities,
for triage in low-access areas) and to smooth referral pathways (Berg et al., 2012; Kim et al., 2024).
4. Equity-focused implementation science.
Evaluate how each pathway (MG+AI, MG+US, AI-POCUS) performs in underserved, rural, and minority
populations, measuring access, follow-up completion, and treatment delays to ensure detection gains translate
into survival gains from breast cancer diangosis(Nicholson et al., 2024; Kim et al., 2024).
5. Standardized reporting
Adopt common data elements (lesion-level labels, density categories, BI-RADS outcomes) and publish
interval cancer rates and node-negative proportions routinely so programs can benchmark beyond CDR (Lee et
al., 2024; Ohuchi et al., 2016).
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1421
www.rsisinternational.org
Figure 10. Workflow framework for breast cancer early detection using Mammography Vs. AI Assisted
POCUS
CONCLUSION
Based on the analysis shared in this manuscript we found that mammography combined with handheld
ultrasound (MG+US) can raise cancer detection rates, but this approach typically lowers specificity and
increases unnecessary downstream benign biopsies. In contrast, when using AI-assisted POCUS showed
promise as an efficient front-line screening strategy.
Consistent with prior published literature on ultrasound screening, AI assisted interpretation was associated
with fewer interval cancers, which is consistent with earlier detection of clinically significant lesions.
Our conclusion, that embedding AI decision support at the bedside enables rapid risk stratification and safe
rule out mechanism of benign findings. This reduces unnecessary specialty referrals and potentially avoidable
biopsies without compromising safety. This real time triage shortens time to diagnostic resolution and
increases throughput, creating opportunities for more cost effective care in settings with limited imaging
capacity.
Collectively, these findings support AI-assisted POCUS as a practical screening modality, especially for
individuals with dense breasts. With standardized image acquisition protocols, robust machine learning
pipelines, and targeted clinician training, AI-assisted POCUS can accelerate diagnostic workflows, expedite
treatment initiation for true positives, and improve patient experience. AI-POCUS holds promise in low-
resource settings, real-time triage, and in facilitating earlier lesion identification where conventional radiology
is limited.
As shared above, future studies should evaluate long-term results, program-level expenses,
and large-scale health system deployment, including prospective model drift monitoring. To reduce operator-
dependent variability, particularly when handheld devices are scaled, we need to develop competency
frameworks, image-acquisition protocols, and recurring proficiency testing. In accordance with established
screening recommendations, it should also assess performance equity across subgroups and integration
mechanisms with transparent and auditable clinical decision support. Additionally, we must establish post-
deployment surveillance for AI: to monitor version upgrades, recall/CDR drift, and bias across subgroups;
standardize calibration and reporting so that outcomes are similar across sites and suppliers.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1422
www.rsisinternational.org
ACKNOWLEDGMENTS
We acknowledge all our collaborators and team members who made this work possible at Validus Institute Inc.
We also acknowledge Portland State University, Department of Electrical and Computer Engineering, led by
Dr. Faryar Etesami for his guidance and support.
REFERENCES
1. Berg, W. A., Zhang, Z., Lehrer, D., Jong, R. A., Pisano, E. D., Barr, R. G., Böhm-Vélez, M., Mahoney,
M. C., Evans, W. P., Larsen, L. H., Morton, M. J., Mendelson, E. B., Farria, D. M., Cormack, J. B.,
Marques, H. S., Adams, A., Yeh, N. M., Gabrielli, G., & ACRIN 6666 Investigators. (2012). Detection
of breast cancer with addition of annual screening ultrasound or a single screening MRI to
mammography in women with elevated breast cancer risk. JAMA, 307(13), 1394–1404.
https://doi.org/10.1001/jama.2012.388
2. Bunnell, A., Valdez, D., Strand, F., Glaser, Y., Sadowski, P., & Shepherd, J. A. (2025). Artificial
intelligence–enhanced handheld breast ultrasound for screening: A systematic review of diagnostic test
accuracy. PLOS Digital Health, 4(9), e0001019. https://doi.org/10.1371/journal.pdig.0001019
3. Duffy, S. W., Agbaje, O., Tabár, L., Vitak, B., Bjurstam, N., Björneld, L., Myles, J. P., & Warwick, J.
(2005). Overdiagnosis and overtreatment of breast cancer: Estimates of overdiagnosis from two trials
of mammographic screening for breast cancer. Breast Cancer Research, 7(6), 258.
https://doi.org/10.1186/bcr1354
4. Eisemann, N., Bunk, S., Mukama, T., Baltus, H., Elsner, S. A., Gomille, T., Hecht, G., Heywang-
Köbrunner, S., Rathmann, R., Siegmann-Luz, K., Töllner, T., Vomweg, T. W., Leibig, C., & Katalinic,
A. (2025). Nationwide real-world implementation of AI for cancer detection in population-based
mammography screening. Nature Medicine. https://doi.org/10.1038/s41591-024-03408-6
5. Giordano, L., von Karsa, L., Tomatis, M., Majek, O., de Wolf, C., Lancucki, L., del Moral, A., &
Esperanza, M. (2012). Mammographic screening programmes in Europe: Organization, coverage and
participation. Journal of Medical Screening, 19(Suppl 1), 72–82.
https://doi.org/10.1258/jms.2012.012085
6. Kim, S., Fischetti, C., Guy, M., Hsu, E., Fox, J., & Young, S. D. (2024). Artificial intelligence (AI)
applications for point-of-care ultrasound (POCUS) in low-resource settings: A scoping review.
Diagnostics, 14(15), 1669. https://doi.org/10.3390/diagnostics14151669
7. Khazaei, H., Khajehee, B., Khazaei, D., Oteibi, M., Abbas, K., Balaguru, B. (2025). Leveraging Vertex
AI for automated ultrasound image analysis: A comprehensive review. IJLTEMAS,14(8), 264–274.
https://doi.org/10.51583/IJLTEMAS.2025.1408000033
8. Lee, S. E., Yoon, J. H., Son, N. H., Han, K., & Moon, H. J. (2024). Screening in patients with dense
breasts: Comparison of mammography, artificial intelligence, and supplementary ultrasound. American
Journal of Roentgenology, 222(1), e2329655. https://doi.org/10.2214/AJR.23.29655
9. Nicholson, W. K., Silverstein, M., Wong, J. B., Barry, M. J., Chelmow, D., Coker, T. R., … U.S.
Preventive Services Task Force. (2024). Screening for breast cancer: U.S. Preventive Services Task
Force recommendation statement. JAMA, 331(22), 1918–1930. https://doi.org/10.1001/jama.2024.5534
10. Ohuchi, N., Suzuki, A., Sobue, T., Kawai, M., Yamamoto, S., Zheng, Y. F., … Ishida, T. (2016).
Sensitivity and specificity of mammography and adjunctive ultrasonography to screen for breast cancer
in the Japan Strategic Anti-cancer Randomized Trial (J-START): A randomized controlled trial. The
Lancet, 387(10016), 341–348. https://doi.org/10.1016/S0140-6736(15)00774-6
11. Oteibi, M., Tamimi, A., Abbas, K., Tamimi, G., Khazaei, D., & Khazaei, H. (2024). Advancing digital
health using AI and machine learning solutions for early ultrasonic detection of breast disorders in
women. International Journal of Research and Innovation in Applied Science, 11(9), 590–601.
https://doi.org/10.51244/IJRSI.2024.11110039
12. Oteibi, M., Khazaei, H., Abbas, K., Balaguru, B., Williams, A. R., & Etesami, F. (2025). Breast
imaging and omics for non-invasive integrated classification (BIONIC). International Journal of
Research and Innovation in Applied Science, 10(8), 826–835.
https://doi.org/10.51584/IJRIAS.2025.100800094
13. Oteibi, M., Tamimi, A., Abbas, K., Tamimi, G., Khazaei, D., & Khazaei, H. (2025). Breast tumor
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
Page 1423
www.rsisinternational.org
ultrasound: Clinical applications, diagnostic features, and integration with AI. International Journal of
Research and Innovation in Applied Science, 11(11), 518–527.
https://doi.org/10.51584/IJRIAS.2025.100800034
14. Paci, E. (2012). Summary of the evidence of breast cancer service screening outcomes in Europe and
first estimate of the benefit and harm balance sheet. Journal of Medical Screening, 19(Suppl 1), 5–13.
https://doi.org/10.1258/jms.2012.012077
15. Şenkaya, A. R., Arı, S. A., Karaca, İ., Kebapçı, E., Öztekin, D. C., & et al. (2023). Association of
polycystic ovary syndrome with mammographic density in Turkish women: A population-based case-
control study. Revista da Associação Médica Brasileira, 69(10), e20230138.
https://doi.org/10.1590/1806-9282.20230138