Discover One Intelligent Experience at RSNA 2026

How closely do radiologists' final assessments align with AI-generated thyroid ultrasound reports in routine clinical practice? This large-scale, real-world study evaluated radiologist adoption of AI-generated thyroid nodule assessments across 297 clinics in the United States. Researchers compared AI-generated structured reports with final radiologist-signed reports, examining agreement across the five TI-RADS descriptors and assigned risk categories.

The analysis included 95,464 studies and 187,292 thyroid nodules. Radiologists retained all AI-generated TI-RADS descriptors without modification in 88.7% of nodules, with agreement across individual descriptors ranging from 94.7% to 97.9%. The findings demonstrate high concordance between AI-generated assessments and radiologists' final reports, supporting the potential for greater standardization in thyroid ultrasound reporting.

The research was conducted by Sadaf Monajemi, Leeann D. Louis, Annie Ng, Melih Engin, Shruti Surendrakumar, Sebastian Parkitny, Jamie Vo, Milad Mohammadzadeh, Benjamin Wildman-Torbiner, and Bryan Haslam.

Interested in the real-world evidence behind AI-assisted thyroid ultrasound reporting? Explore the study's methodology and findings during this RSNA 2026 scientific session.

For attending this presentation, please register here.

Breast arterial calcifications (BAC) detected on routine mammography are recognized markers of cardiovascular and peripheral vascular disease, yet their detection and reporting remain inconsistent. This retrospective, blinded, multicenter study evaluated the standalone performance of an AI algorithm for BAC detection and localization across full-field digital mammography (FFDM) and digital breast tomosynthesis (DBT), examining its consistency across breast densities and patient age groups. The analysis included 1,775 screening mammography examinations from 37 U.S. clinical sites, with independent expert radiologist review establishing the reference standard.

The AI demonstrated 96.0% sensitivity on FFDM and 90.5% on DBT, with areas under the curve (AUC) of 0.979 and 0.965, respectively. Performance was also evaluated across breast density and age subgroups, with findings supporting the potential for more standardized, objective BAC detection and reporting in routine mammography.

The research was conducted by Akira Hasegawa, James Pike, Nikolaos Gkanatsios, Jonathan Go, Brittaney Everitt, Jun Ge, Yinhao Ren, Annie Ng, Jiye Kim, and Bryan Haslam.

Discover the study's methodology and findings, including how AI performance compares across mammography modalities and patient subgroups, during this RSNA 2026 scientific poster discussion.

For attending this presentation, please register here.

Can deep-learning models improve lung cancer risk prediction from low-dose CT (LDCT) screening examinations? This retrospective study evaluated the performance of the image-based Kiraly model against the established Brock risk prediction model and Sybil, another deep-learning model. Researchers analyzed 1,191 LDCT examinations from four U.S. imaging centers, including 195 participants diagnosed with lung cancer, to assess the models' ability to distinguish malignant from benign findings in an independent screening population.

The Kiraly model demonstrated higher discriminatory performance, with an area under the curve (AUC) of 0.93 compared with 0.89 for Sybil. Among Lung-RADS 3+ examinations, Kiraly achieved an AUC of 0.89, compared with 0.79 for Brock and 0.84 for Sybil. These findings suggest that image-based deep learning could contribute to more accurate lung cancer risk stratification in LDCT screening, supporting further evaluation of its potential clinical utility.

The research was conducted by Maurits Engbersen, Marios Koulakis, Jeroen Mollink, Atilla P. Kiraly, PhD, Bryan Haslam, PhD

Discover how these AI models compare in predicting lung cancer risk and explore the implications for LDCT screening during this RSNA 2026 scientific presentation.

For attending this presentation, please register here.

Can AI identify early signs of breast cancer on screening mammograms years before diagnosis? This retrospective analysis of the ASSURE cohort investigated whether a commercially available AI algorithm could flag the breast subsequently diagnosed with cancer on prior digital breast tomosynthesis (DBT) examinations. Researchers evaluated 1,361 women with at least one prior DBT examination, drawn from a cohort of 2,762 women with screen-detected breast cancer across 109 U.S. sites. AI findings in the cancer-affected breast were compared with the unaffected breast, with analyses stratified by race and ethnicity, breast density, and time before diagnosis.

On the most recent prior screening examination, AI flagged the cancer-affected breast in 26.5% of women, compared with 5.1% for the unaffected breast. Elevated flag rates were observed across patient subgroups and remained detectable up to four years before diagnosis. These findings provide evidence of AI-detectable signals preceding breast cancer diagnosis in a diverse U.S. screening population and support further investigation into AI's potential role in earlier cancer detection.

The research was conducted by Leeann D. Louis, Viviane Bui, Annie Ng, Diana S. M. Buist, Christoph I. Lee, and Bryan Haslam.

What can prior mammograms reveal about breast cancer before diagnosis? Explore the study's methodology and findings during this late-breaking scientific presentation at RSNA 2026.

For attending this presentation, please register here.

Can AI improve prostate cancer detection and reduce variability in MRI interpretation across diverse patient populations, imaging conditions, and levels of radiologist experience? This blinded, multi-reader, multi-case study evaluated an AI platform providing automated lesion segmentation and lesion-level risk classification. Twelve board-certified radiologists interpreted 247 prostate MRI examinations with and without AI assistance, allowing researchers to assess differences in cancer detection performance and lesion localization across clinically relevant subgroups.

AI-assisted interpretation demonstrated a 15.7% relative improvement in prostate cancer detection performance, with mean AUC-LROC increasing from 0.626 to 0.724. Improvements were observed for clinically significant prostate cancer and across patient demographic groups, MRI scanner vendors, field strengths, and reader experience levels. AI assistance also increased inter-reader agreement in lesion segmentation, supporting the potential for more consistent lesion localization and integration with fusion biopsy workflows.

The research was conducted by Brittaney Everitt, Valerio Fortunati, Jiye Kim, Jim Tol, Abigail Ko, Guus Kolpa, Felix Leufkens, and Bryan Haslam.

Explore the study's findings on AI-assisted prostate MRI interpretation, including its performance across diverse clinical subgroups, during this RSNA 2026 scientific presentation.

For attending this presentation, please register here.

Can AI-driven breast cancer screening improve the detection of aggressive cancers in women with dense breasts? This large-scale analysis from the ASSURE study evaluated the clinical impact of a multistage AI-driven screening workflow across breast density groups and cancer subtypes. Conducted across 109 imaging sites in four U.S. states, the study compared standard-of-care screening with an AI-assisted workflow combining concurrent AI interpretation and specialist safeguard review. Researchers assessed cancer detection rates, positive predictive values, and the detection of aggressive versus non-aggressive cancers in women with dense and non-dense breasts.

The AI-driven workflow demonstrated higher cancer detection rates and positive predictive values across breast density groups. In women with dense breasts, detection of aggressive cancers increased significantly, with a relative rate of 1.68 (95% CI: 1.17–2.40; p=0.005) compared with standard-of-care screening. These findings suggest that AI-assisted screening workflows may improve the detection of clinically relevant cancers, particularly aggressive subtypes in dense breast tissue.

The research was conducted by Annie Ng, Leeann D. Louis, Edgar A. Wakelin, Matthew McCabe, Jiye Kim, Christoph Lee, Diana S. M. Buist, and Bryan Haslam.

How does breast density influence the clinical impact of AI-assisted cancer screening? Join this RSNA 2026 scientific presentation to explore the ASSURE study's findings and their implications for detecting aggressive breast cancers.

For attending this presentation, please register here.

Can AI enable more accurate and equitable breast cancer risk prediction using digital breast tomosynthesis (DBT) alone? This retrospective study evaluated an image-only deep-learning model designed to predict five-year breast cancer risk without using race as an input. Researchers analyzed 65,650 women undergoing DBT screening, including 1,040 breast cancer cases, with at least five years of follow-up. The study assessed model discrimination and calibration, including performance for invasive breast cancer and across White and Black patient subgroups.

The DBT-based model achieved an overall five-year AUC of 0.693, compared with a reported AUC of 0.66 for a previously published model based on full-field digital mammography (FFDM). Performance remained comparable for invasive cancers (AUC 0.692), with AUCs of 0.686 in White women and 0.720 in Black women. The findings support the potential of image-based AI to inform more personalized breast cancer risk assessment while maintaining consistent performance across the racial subgroups evaluated.

The research was conducted by James Pike, Ross Lawrence, Hyunkwang Lee, Brittany Everitt, Lee Tirrell, Jiye Kim, and Bryan Haslam.

Explore how image-only deep learning could contribute to more personalized, risk-based breast cancer screening, and learn more about the study's performance across patient populations during this RSNA 2026 scientific presentation.

For attending this presentation, please register here.

Can incorporating prior mammograms reduce false-positive AI detections while maintaining cancer detection performance? This retrospective, multicenter study evaluated whether adding prior digital breast tomosynthesis (DBT) examinations to AI analysis improves the specificity of breast cancer detection. Researchers analyzed 684 screening DBT examinations, including 180 cancer cases and 504 non-cancer cases, from seven U.S. clinical sites. Each examination was evaluated with and without a prior screening study acquired 12–36 months earlier, allowing researchers to assess the impact of longitudinal imaging information on AI detection and localization performance.

Incorporating prior examinations resulted in a 30% reduction in false positives per image, from 0.505 to 0.351, while maintaining comparable cancer detection and localization performance. A 16% reduction was also observed for intermediate- and high-suspicion findings combined. These results suggest that incorporating temporal context into AI analysis can help distinguish stable findings from potentially significant changes, reducing unnecessary AI markings and supporting more focused radiologist review.

The research was conducted by Brittaney Everitt, Jim Pike, Ross Lawrence, Hyunkwang Lee, Jiye Kim, and Bryan Haslam.

Explore how incorporating prior mammograms into AI analysis could support more efficient breast cancer screening workflows. Join this RSNA 2026 scientific poster discussion to learn more about the study's methodology, findings, and potential clinical implications.

For attending this presentation, please register here.