AI improves breast cancer detection rates while reducing burden on radiologists

AI-powered tools in mammography screening could lead to breakthrough improvements in cancer detection, helping radiologists catch more cancers earlier while reducing unnecessary patient recalls. Helpful.

study: National real-world implementation of AI for cancer detection in population-based mammography screening. Image credit: Gorodenkoff / Shutterstock

In a recent study published in the journal natural medicineresearchers investigated the impact of artificial intelligence (AI) on cancer detection and recall rates.

Mammography screening contributes to reducing breast cancer-related mortality. Additionally, improved sensitivity and specificity of screening may reduce interval cancer rates and recall rates, making treatment more effective for breast cancer patients. Screening programs generate large numbers of mammograms, and most require interpretation by two radiologists.

Additionally, consensus conferences may be required to achieve high specificity and sensitivity. As a result, a radiologist's job involves the repetitive task of interpreting a large number of images each week. This workload is likely to increase, especially as recent guidelines recommend mammography screening for more age groups. Incorporating AI into cancer screening programs could alleviate some of the problems.

Studies show that AI is as accurate as, and in some cases better than, radiologists. Several studies have observed increased cancer detection with AI-integrated workflows, despite inconsistent results regarding recall. Nevertheless, the authors of this study note that the small samples and low heterogeneity of radiologists, screening facilities, and equipment vendors in these early studies limit generalizability. I emphasized.

Research and findings

In this study, researchers evaluated the impact of AI on cancer recall and detection rates. The research was conducted within the following scope: Breast cancer screening A German program aimed at asymptomatic people aged 50 to 69. Data was collected from multiple examination venues that installed AI systems from July 2021 to February 2023.

In the screening program, four mammograms were obtained for each participant and initially read by two independent radiologists. If one of the radiologists deemed the case suspicious, a consensus conference was held. If suspicious findings persist during the meeting, participants will be recalled for further evaluation.

Examinations whose reports were read and submitted by at least one radiologist using an AI-enabled viewer were included in the AI ​​group. Exams that were not submitted using the AI-based viewer were included in the control group. Radiologists can use existing (non-AI-based) software or AI-supported viewers.

The AI ​​system Vara MG leveraged two key features. One is regular triage, which flags highly suspicious tests as normal, and the other is safety net, which highlights highly suspicious cases and locates suspicious areas. This safety net was intended to reduce missed diagnoses by having radiologists review suspicious findings flagged by AI.

A total of 461,818 women who underwent mammography were included, and 119 radiologists interpreted the exams. Of these, 260,739 were included in the AI ​​group and 201,079 were included in the control group. Approximately 42 per 1,000 women had suspicious findings and were recalled for further testing. About a quarter of them had a biopsy, or more than 6 per 1,000 women. diagnosed with breast cancer.

The AI ​​system classified 59.4% of the exams as normal, significantly reducing the workload of radiologists. Safety nets were triggered for 1.5% of tests in the AI ​​group, leading to 541 recalls and 208 cancer diagnoses. In addition, 3.1% of tests in the AI ​​group that were determined to be normal by the AI ​​underwent further evaluation by the consensus group, resulting in an additional 20 cancer diagnoses. Breast cancer detection rate (BCDR) was 6.7 and 5.7 per 1,000 women in the AI ​​and control groups, respectively.

The AI ​​group showed statistically higher BCDR and slightly lower recall than the control group. The positive predictive value (PPV) of recall for the AI ​​and control groups was 17.9% and 14.9%, respectively. AI group was 8.2% higher. biopsy higher than the control group. Nevertheless, the AI ​​group had a higher biopsy PPV (64.5%) than the control group (59.2%).

Wider implications and future considerations

This study highlighted that integrating AI into screening workflows has the potential to increase the detection of ductal carcinoma in situ (DCIS) cases. Although this may indicate early detection, these cases do not necessarily progress to invasive cancer, raising concerns about overdiagnosis and overtreatment of DCIS. The long-term impact on interval cancer incidence and stage distribution requires further follow-up over 2-3 years.

Additionally, the researchers noted that rejected safety net cases may be a missed opportunity to detect cancer early or demonstrate value in reducing unnecessary recalls, making further analysis more likely. He emphasized that this is an important area that needs to be addressed.

Overall, the AI ​​approach to mammography screening provided confident doubt predictions and confident normal predictions. The BCDR of the AI ​​group was 17.6% higher than that of the control group. Using AI also slightly decreased recall, although it was not statistically significant. These findings contribute to the evidence base that AI-assisted mammography screening is safe, feasible, and can reduce workload.