Keyword: Radiographic Image Interpretation
2 results found.
Congress Abstract
Oncology, Nuclear Medicine and Transplantology, 2(3, Suppl. 1), 2026, onmt_A24, https://doi.org/10.63946/onmt/19313
ABSTRACT:
Background: Dense breast tissue reduces mammographic sensitivity and raises breast cancer risk. Supplemental imaging and artificial intelligence supported mammography reading have been tested prospectively. The European Society of Breast Imaging recommends supplemental magnetic resonance imaging for extremely dense breasts, whereas the American College of Physicians in 2026 advised against it for average risk women with dense breasts. This review summarises what the principal studies measured and found.
Objective: To review the published evidence on both strategies in dense breasts, reporting each study's endpoints and results.
Materials and Methods: Narrative review of primary publications, published online 2019 to 2025, quoted verbatim. Four large prospective multicentre studies reporting on dense breasts were selected, two per strategy: the randomised trials DENSE (Netherlands, magnetic resonance imaging) and BRAID (United Kingdom, abbreviated magnetic resonance imaging, automated ultrasound and contrast-enhanced mammography), and the observational studies PRAIM (Germany, artificial intelligence supported double reading) and ASSURE (United States, artificial intelligence supported single reading with safeguard review).
Results: DENSE randomised 40,373 women with extremely dense breasts and normal mammography to invitation for supplemental magnetic resonance imaging or to mammography alone: interval cancers 2.5 versus 5.0 per 1000; among the 59% accepting, detection was 16.5 and false positives 79.8 per 1000. BRAID randomised 9361 women with dense breasts and a negative mammogram: detection was 17.4 per 1000 examinations with abbreviated magnetic resonance imaging, 19.2 with contrast-enhanced mammography and 4.2 with automated ultrasound, the contrast-based modalities not differing significantly; recall 9.7%, 9.7% and 4.0%, median invasive size 10, 11 and 22 millimetres. In PRAIM (463,094 women), artificial intelligence supported double reading was associated with detection of 6.7 versus 5.7 per 1000 (17.6% higher) and non-inferior recall; in dense breasts the 18.7% increase was not statistically significant. In ASSURE (579,583 tomosynthesis examinations, single reading), detection with the artificial intelligence workflow was 5.6 versus 4.6 per 1000 (21.6% higher), recall 11.1% versus 10.6%, and detection in dense breasts 22.7% higher.
Conclusions: In BRAID, abbreviated magnetic resonance imaging and contrast-enhanced mammography detected three times as many invasive cancers as automated ultrasound, at half the size with more than twice the recall, and did not differ significantly from each other. In DENSE, invitation to supplemental magnetic resonance imaging halved the interval cancer rate. Artificial intelligence support was associated with higher detection in PRAIM without higher recall and in ASSURE with slightly higher recall; neither was randomised. No included study measured breast cancer mortality; survival benefit remains undemonstrated and overdiagnosis unquantified.
Objective: To review the published evidence on both strategies in dense breasts, reporting each study's endpoints and results.
Materials and Methods: Narrative review of primary publications, published online 2019 to 2025, quoted verbatim. Four large prospective multicentre studies reporting on dense breasts were selected, two per strategy: the randomised trials DENSE (Netherlands, magnetic resonance imaging) and BRAID (United Kingdom, abbreviated magnetic resonance imaging, automated ultrasound and contrast-enhanced mammography), and the observational studies PRAIM (Germany, artificial intelligence supported double reading) and ASSURE (United States, artificial intelligence supported single reading with safeguard review).
Results: DENSE randomised 40,373 women with extremely dense breasts and normal mammography to invitation for supplemental magnetic resonance imaging or to mammography alone: interval cancers 2.5 versus 5.0 per 1000; among the 59% accepting, detection was 16.5 and false positives 79.8 per 1000. BRAID randomised 9361 women with dense breasts and a negative mammogram: detection was 17.4 per 1000 examinations with abbreviated magnetic resonance imaging, 19.2 with contrast-enhanced mammography and 4.2 with automated ultrasound, the contrast-based modalities not differing significantly; recall 9.7%, 9.7% and 4.0%, median invasive size 10, 11 and 22 millimetres. In PRAIM (463,094 women), artificial intelligence supported double reading was associated with detection of 6.7 versus 5.7 per 1000 (17.6% higher) and non-inferior recall; in dense breasts the 18.7% increase was not statistically significant. In ASSURE (579,583 tomosynthesis examinations, single reading), detection with the artificial intelligence workflow was 5.6 versus 4.6 per 1000 (21.6% higher), recall 11.1% versus 10.6%, and detection in dense breasts 22.7% higher.
Conclusions: In BRAID, abbreviated magnetic resonance imaging and contrast-enhanced mammography detected three times as many invasive cancers as automated ultrasound, at half the size with more than twice the recall, and did not differ significantly from each other. In DENSE, invitation to supplemental magnetic resonance imaging halved the interval cancer rate. Artificial intelligence support was associated with higher detection in PRAIM without higher recall and in ASSURE with slightly higher recall; neither was randomised. No included study measured breast cancer mortality; survival benefit remains undemonstrated and overdiagnosis unquantified.
Review Article
Oncology, Nuclear Medicine and Transplantology, 2(3, Suppl. 1), 2026, onmt_A1, https://doi.org/10.63946/onmt/19254
ABSTRACT:
Background: Lung cancer leads cancer mortality in Kazakhstan: the International Agency for Research on Cancer estimates 2,798 new cases and 2,617 deaths for 2024. Randomised trials show that low-dose computed tomography screening lowers lung cancer mortality, but population screening brings heavy reading workloads and many false positives. This review examines the published evidence on artificial intelligence in this setting.
Objective: To review published evidence on artificial intelligence in low-dose computed tomography lung cancer screening, with implications for programmes in Kazakhstan and Central Asia.
Materials and Methods: Narrative review of peer-reviewed publications, 2011 to August 2026, in PubMed and publisher databases. Screening trials required a mortality endpoint; artificial intelligence studies required histological outcomes, an expert panel reference standard or randomisation. Included: the National Lung Screening Trial (2011), the Dutch-Belgian screening trial (2020), Sybil (2023), the United Kingdom Lung Cancer Screening trial validation (2025), the 4-IN-THE-LUNG-RUN feasibility study (2025) and a prospective single-centre randomised trial (2026).
Results: The National Lung Screening Trial reduced lung cancer mortality by 20.0 percent (95 percent confidence interval 6.8 to 26.7) versus chest radiography; 96.4 percent of positive screens were false positives. The Dutch-Belgian trial reported a ten-year lung cancer mortality rate ratio of 0.76 (95 percent confidence interval 0.61 to 0.94) among male participants versus no screening. Neither trial used artificial intelligence. Sybil predicted one-year cancer risk from one scan with areas under the receiver operating characteristic curve of 0.92, 0.86 and 0.94 in three retrospective cohorts. In 1,252 United Kingdom baseline scans, an artificial intelligence first reader detected all 31 histologically confirmed cancers, one below its volume threshold (negative predictive value 99.8 percent), with an estimated maximum workload reduction of 79 percent. In 3,678 European baseline scans, artificial intelligence negative misclassifications were 0.8 percent against 11.1 percent for radiologists; its positive misclassifications were 5.7 percent against 0.5 percent for radiologists. In a randomised trial in asymptomatic individuals, artificial intelligence assistance raised detection of Lung Imaging Reporting and Data System positive nodules from 10.3 to 16.9 percent with no significant change in interpretation time.
Conclusions: The mortality benefit belongs to low-dose computed tomography screening itself; no included artificial intelligence study measured mortality. Evidence is strongest for artificial intelligence as a first reader ruling out negative baseline scans while radiologists read the rest; detection assistance raises nodule yield and positive misclassifications; single-scan risk prediction remains retrospective. For Kazakhstan and Central Asia these applications address the reading capacity and false-positive burden that limit programme feasibility, provided tools are validated locally against histological outcomes.
Keywords: Lung Neoplasms; Early Detection of Cancer; Tomography, X-Ray Computed; Artificial Intelligence; Radiographic Image Interpretation, Computer-Assisted; Kazakhstan
Objective: To review published evidence on artificial intelligence in low-dose computed tomography lung cancer screening, with implications for programmes in Kazakhstan and Central Asia.
Materials and Methods: Narrative review of peer-reviewed publications, 2011 to August 2026, in PubMed and publisher databases. Screening trials required a mortality endpoint; artificial intelligence studies required histological outcomes, an expert panel reference standard or randomisation. Included: the National Lung Screening Trial (2011), the Dutch-Belgian screening trial (2020), Sybil (2023), the United Kingdom Lung Cancer Screening trial validation (2025), the 4-IN-THE-LUNG-RUN feasibility study (2025) and a prospective single-centre randomised trial (2026).
Results: The National Lung Screening Trial reduced lung cancer mortality by 20.0 percent (95 percent confidence interval 6.8 to 26.7) versus chest radiography; 96.4 percent of positive screens were false positives. The Dutch-Belgian trial reported a ten-year lung cancer mortality rate ratio of 0.76 (95 percent confidence interval 0.61 to 0.94) among male participants versus no screening. Neither trial used artificial intelligence. Sybil predicted one-year cancer risk from one scan with areas under the receiver operating characteristic curve of 0.92, 0.86 and 0.94 in three retrospective cohorts. In 1,252 United Kingdom baseline scans, an artificial intelligence first reader detected all 31 histologically confirmed cancers, one below its volume threshold (negative predictive value 99.8 percent), with an estimated maximum workload reduction of 79 percent. In 3,678 European baseline scans, artificial intelligence negative misclassifications were 0.8 percent against 11.1 percent for radiologists; its positive misclassifications were 5.7 percent against 0.5 percent for radiologists. In a randomised trial in asymptomatic individuals, artificial intelligence assistance raised detection of Lung Imaging Reporting and Data System positive nodules from 10.3 to 16.9 percent with no significant change in interpretation time.
Conclusions: The mortality benefit belongs to low-dose computed tomography screening itself; no included artificial intelligence study measured mortality. Evidence is strongest for artificial intelligence as a first reader ruling out negative baseline scans while radiologists read the rest; detection assistance raises nodule yield and positive misclassifications; single-scan risk prediction remains retrospective. For Kazakhstan and Central Asia these applications address the reading capacity and false-positive burden that limit programme feasibility, provided tools are validated locally against histological outcomes.
Keywords: Lung Neoplasms; Early Detection of Cancer; Tomography, X-Ray Computed; Artificial Intelligence; Radiographic Image Interpretation, Computer-Assisted; Kazakhstan