Oncology researchers have been exploring whether artificial intelligence can be used to detect cancer faster or more accurately than radiologists. A promising new breast cancer study by German researchers suggests that AI could detect breast cancer in ways that radiologists may easily miss. However, other research has indicated that AI sometimes places too much importance on irrelevant mammogram details.
Let’s look closer at how AI could potentially help improve cancer diagnoses and what details still need to be ironed out before this technology can be used to make critical cancer treatment decisions.
The Importance of Early Detection in Cancer Treatment
The stage at which a cancer is detected significantly influences expected outcomes for the patient. Early detected cancers are generally much easier to treat, which means patients whose cancers are discovered early are much more likely to survive. Early-stage cancers are usually localized within one part of the body, while late-stage cancers often have spread to other parts.
Common cancers like breast cancer and prostate cancer can be detected earlier through regular screenings. Mammograms, which can detect breast cancer before symptoms begin, are generally recommended for women 40-45 years old.
However, mammogram screening is still flawed, and false positives are common. A promising new German study published in the European Journal of Radiology showed that AI could potentially be used to make early diagnoses and reduce false positives.
German Study Finds AI Could Detect Breast Cancers Missed by Radiologists
According to the findings of the peer-reviewed European Journal of Radiology study, AI can be used to accurately detect interval breast cancers, which show up between regularly scheduled mammograms. In the United States, doctors generally recommend annual screenings for women between ages 40 and 54, while those over age 55 can get a mammogram every two years (with the choice to continue with annual screenings).
The study showed that only about 16 percent of interval cancers are likely visible during previous screenings, while about 20 percent might be undetectable by the human eye. These minimal signs are often missed by radiologists. The findings of this study show that AI could be used to detect these commonly missed interval cancers.
The Mammography Reference Centre North in Oldenburg, Germany, partnered with German deep-tech start-up Vara to conduct this study. Vara analyzed 2,396 mammogram screens from women who were later diagnosed with interval breast cancer.
The analysis determined that AI could accurately detect and localize 27.5 percent of false-negative cases and over 12 percent of minimal signs of breast cancer. In 3 percent of cases, the AI algorithm accurately located where breast cancer would later appear.
The algorithm classified the mammograms into three categories: cancer suspicious, normal, and unconfident. Cancer suspicious and unconfident mammograms were forwarded to the radiologist. To avoid bias, the radiologist was not told which samples might have cancer. If the radiologist’s prediction did not match the algorithm’s, the algorithm would raise an alarm.
Issues with AI Cancer Diagnoses
While the potential cancer treatment applications of AI are exciting, this technology will not put radiologists out of work anytime soon. AI systems and radiologists show major differences in breast cancer screenings, including some AI errors not made by radiologists. According to another recent study published in the journal Nature Scientific Reports, both human and AI methods should be used together to make medical diagnoses.
This study focused on an AI system called deep neural networks, which consist of computer-simulated neurons. These neurons can be programmed to “learn” by constructing layers of computing elements and determining how to make calculations based on data input. This process is known as deep learning.
The researchers compared breast cancer screenings made by deep neural networks with those read by radiologists. They found that the two methods differed significantly in their diagnoses of soft-tissue lesions.
AI systems were shown to consider minor mammogram details that radiologists consider irrelevant. According to the authorities, these discrepancies in readings must be understood and addressed before oncologists can trust AI systems to assist in vital cancer treatment decisions.
How Radiologists Can Use AI to Make Better Treatment Decisions
Based on the findings of the Nature Scientific Reports study, the authors believe that the combination of AI and radiologist readings has the potential to improve cancer screenings. Radiologists should view these AI algorithms as tools that can help them quickly detect breast cancers that they might otherwise miss on their own.
Radiologists primarily rely on brightness and shape to identify tumors, while deep neural networks analyze tiny details spread throughout mammogram images. Many of these details are concentrated in regions that most radiologists consider unimportant.
This study helps oncologists better understand the differences between human and AI perception in medical diagnosis, which brings researchers closer to transitioning AI from academic study to clinical practice.