AI-Supported Mammography Is As Good As Two Breast Radiologists Working Together: Study In Lancet

AI-supported mammography: The study, published August 2 in The Lancet Oncology journal, says that AI-based mammography can halve the workload of radiologists trying to detect breast cancer. 

AI-supported mammography: While standard double reading without AI detects five cancers per 1,000 women, AI-supported screening helped detect six cancers per 1,000 screened women. This means that AI helped detect one additional cancer for every 1,000 women screened. 

Source : Getty/ABP Live

Artificial intelligence (AI)-supported mammography analysis is as good as two breast radiologists working together, and halves the screen-reading workload, according to a new study. As part of the research, a safety analysis of a randomised controlled trial involving more than 80,000 Swedish women was conducted. The study, published August 2 in The Lancet Oncology journal, says that AI-based mammography can halve the workload of radiologists trying to detect breast cancer. 

The study also states that certain trial results are not expected for several years. These are the results that will show whether using AI to interpret mammography images will help reduce interval cancers in 100,000 women followed over two years, and also whether the use of AI in mammography screening is vindicated. Interval cancers are cancers detected after a normal mammogram screening, and before the next scheduled mammogram screening. 

Since breast cancer screening through mammography helps detect breast cancer at an earlier, more treatable stage, the technique has been found to improve prognosis and reduce mortality. But, it is estimated that 20 to 30 per cent of interval cancers that should have been detected at the preceding screening mammogram are missed, and suspicious findings often turn out to be benign, or non-cancerous. 

Therefore, two radiologists are required to screen mammograms to ensure high sensitivity, according to European guidelines. This is to ensure that the disease is correctly identified. However, breast radiologists in many countries are scarce, and it takes more than a decade to train a radiologist capable of interpreting mammograms. The countries in which there has been a scarcity of radiologists are the United Kingdom and Sweden. 

In a Lancet statement, Dr Kristina Lång, the lead author on the paper, said promising interim safety results like the one published in The Lancet Oncology journal should be used to inform new trials and programme-based evaluations to address the pronounced radiologist shortage in many countries. She also said that these trials are not enough on their own to confirm that AI is ready to be implemented in mammography screening. This is because researchers still have to understand the implications of incorporating AI in mammography screening on the outcomes of patients, and whether this technique can help detect interval cancers often missed by traditional screening, and prove to be cost-effective. 

How AI-based screening can prove to be effective in breast cancer imaging

The study proposes AI as an automated second reader for mammograms to help reduce the workload of radiologists and improve the accuracy of breast cancer screening. AI-based mammography has been shown to be helpful because it provides radiologists with computer-aided detection marks which highlight suspicious features to reduce false negative results. However, scientists are yet to obtain robust evidence from such trials. 

As many as 80,033 women aged 40 to 80 years were screened between April 2021 and July 2022. All these women had undergone mammogram screening at four sites in southwest Sweden. At these sites, some mammograms were analysed by a commercially available AI-supported mammogram reading system, before being read by one or two radiologists, while others underwent standard analysis, as part of which two radiologists checked the mammograms without AI. 

The radiologists who checked the mammograms after the AI analysis served as the intervention arm, while the radiologists who performed the standard analysis without an AI served as the control arm. 

As part of the analysis, the researchers compared early screening performance with the screen-reading workload in the intervention and control arms. Early screening performance includes cancer detection and false positives. 

The intervention group set the lowest acceptable limit for clinical safety at a cancer detection rate above three cancers per 1,000 screened women.

As part of the analysis, the AI system first analysed the mammography image. After this, the system predicted the risk of cancer on a scale of one to 10, where one represents the lowest risk, and 10 the highest. A radiologist further analysed the mammography image if the risk score was less than 10. When the AI system predicted a risk score of 10, two radiologists analysed the image. 

In 0.8 per cent of cases referred to standard care, or double reading, AI failed to provide a risk score. 

In the case of AI-supported screening, the average recall rates were 2.2 per cent, while in standard double reading without AI, the recall rates were two per cent. 

Of the women recalled from AI-supported screening, 28 per cent had cancer. Meanwhile, of the women recalled from standard screening, 25 per cent had cancer. As many as 244 women were recalled from AI-supported screening, and 203 women were recalled from standard screening. This shows that 41 more cancers were detected with the help of AI. 

While standard double reading without AI detects five cancers per 1,000 women, AI-supported screening helped detect six cancers per 1,000 screened women. This means that AI helped detect one additional cancer for every 1,000 women screened. 

In the AI-supported group, there were 36,886 fewer screen readings by radiologists, compared to the control group. There were 46,345 screen readings by radiologists in the AI-supported group, and 83,231 readings by radiologists in the control group, showing that there was 44 per cent reduction in the screen-reading workload. 

Limitations to the study

The authors noted some limitations to the study, which include the fact that the analysis was conducted at a single centre and was limited to one type of mammography device and one AI system. Other limitations are that technical factors are likely to affect the performance of the system.

However, the authors believe that these factors are likely to be less important than the experience of radiologists, and since the final decision on whether to recall women will be made by radiologists, the outcome is dependent on the performance of the radiologists.

Check ABP Live's stories explaining the science behind various health phenomena, and the articles appearing in the weekly health column here.