AI-assisted breast cancer screening may improve health outcomes and system efficiency compared with conventional double reading.A study from Singapore shows the cost-effectiveness of integrating artificial intelligence (AI) into breast cancer screening, with both AI-assisted and AI-standalone models having an edge over conventional double reading.
“The AI-assisted model delivers cost savings and health gains while retaining clinical oversight, while standalone AI provides greater health gains but at higher costs and false-positive rates,” the investigators said.
Three screening strategies were evaluated: conventional mammography with radiologist double reading, AI-assisted mammography, and AI-standalone interpretation. [Cancers (Basel) 2026;18:836]
With double reading, two radiologists independently interpreted each mammogram; discordant cases were resolved by a third reader. With the AI-assisted strategy, a radiologist did the initial read, another radiologist was replaced by AI (FxMammo), and discordant cases were arbitrated by a third radiologist. Of note, decision-making remained with the radiologist, and AI only served as a supportive tool. With AI only, the AI directly interpreted the result, leaving no clinical autonomy to human readers.
Conventional screening yielded 343 true positives, 144 false negatives, 1,761 false positives, 1,201 undiagnosed cases, and 227 early-stage detections.
Conventional screening cost USD 19.18 million with 218,460.4 QALYs*, the AI-assisted strategy cost USD 18.86 million with 218,476 QALYs, and the AI-only strategy cost USD 20.53 million with 218,532 QALYs.
AI + radiologist
The hybrid strategy showed improved diagnostic performance, with fewer false negatives (n=131), false positives (n=1,225), and undiagnosed cases (n=1,191) and more early-stage detections (n=231).
Moreover, this strategy had the most favourable economic profile, yielding total savings of SGD 316,090 and gaining 15.9 QALYs, resulting in an ICER** of SGD 19,846 cost savings per QALY gained, compared with conventional screening. “This made it the most cost-effective option across low-to-intermediate willingness-to-pay thresholds,” the researchers said.
“Sensitivity analyses identified diagnostic specificity, screening costs, and early-stage detection as key drivers of economic value,” they added.
AI alone
The AI-alone model yielded the largest health gains, delivering more true positives (n=372), fewer undiagnosed cases (n=1,155), and the fewest false negatives (n=85), reflecting higher sensitivity.
AI alone yielded an additional 72 QALYs, with an ICER of SGD 18,743 per QALY gained.
However, this came at the expense of the largest number of false positives (n=4,097), reflecting reduced specificity and a higher recall burden. Moreover, the increased total costs (+SGD 1.35 million) raise clinical and system-level concerns regarding downstream investigations and resource utilization.
According to the investigators, the substantially higher false-positive rate has direct clinical and ethical implications, considering the well-documented cascading consequences of false positives (eg, additional diagnostic imaging, invasive biopsies, increased healthcare resource utilization, psychological distress, anxiety).
“In this context, the elevated false-positive burden represents not only an economic cost but also a key practical and ethical barrier to real-world adoption of fully automated screening strategies,” the researchers explained.
Phased implementation
The strategies were applied to 10,000 women aged 50–69 years in Singapore, with a 50-year horizon. “This reflects Singapore’s Healthier SG breast cancer screening strategy, which offers biennial mammography to women aged 50–69,” the researchers noted.
Taken together, the results demonstrate the potential of AI to improve health outcomes and system efficiency in breast cancer screening.
“The AI-assisted strategy appears to provide the most favourable balance between clinical benefit, cost-effectiveness, and operational feasibility without compromising diagnostic performance … [It] represents a more economically viable and pragmatic transitional model for adoption in breast cancer screening programmes,” the researchers said.
Conversely, the fully AI approach may offer greater detection gains but is tied to higher costs and downstream resource implications, warranting further optimization and evaluation before widespread adoption, they continued.
“Overall, the findings support a phased implementation strategy prioritizing AI-assisted workflows as the most practical pathway for near-term population screening programmes,” the researchers added.