Oncology
Discover how AI makes it possible to detect subtle changes in breast tissue that support early detection of tumours and more accurate mammograms.
Enhancing accuracy
Radiologists have long made use of software modules to enhance the accuracy of mammography.
According Dr Hendrik Schoombee, a radiologist at SCP Radiology that provides radiology services to several Mediclinic hospitals, “One of the earliest and most basic examples is computer-aided detection (CAD), which made it easier to detect abnormal shapes against the background of breast tissue.”
Large language models (LLMs) and deep learning have taken this further. “We now have triage-based tools that alert us early to possible issues, so we can investigate them sooner,” he says. These tools are often used for scans such as CTs and MRIs.
Interpretive tools work slightly differently. Dr Schoombee describes them as a radiologist’s “wingman or co-pilot”.
“They will never replace human input, but they can highlight small changes that may be important.”
For example, the tool has been trained on a large set of mammograms and their clinical results. It uses the patterns it has learned to assess new mammograms and gives each one a score. This score shows how likely it is that cancer may be present, helping radiologists make more informed decisions
Dr Schoombee says this is an especially positive development in the case of slow-growing tumours, where subtle changes may develop over several years. AI can help highlight these small interval changes, making them easier for the radiologist to recognise and assess.
Reducing need for second readings
These tools can help reduce the workload for radiologists.
In many European countries, two radiologists usually check every mammogram. This is because early signs of breast cancer can be subtle and easy to miss. A missed cancer can also have serious consequences.
Some European breast screening programmes are now using AI instead of a second radiologist for mammogram checks. This can reduce the need for routine double reading, while still helping doctors detect cancer accurately.
South Africa doesn’t have the resources for every mammogram to be read twice, which makes an additional review by AI even more valuable.
“It’s like having a second pair of eyes, which helps you make a diagnosis with greater confidence,” Dr Schoombee says. This is especially helpful for radiologists who are generalists, or new to the field, since AI can help to build confidence while they develop their experience.
The benefits for patients are just as great. “The earlier cancer is detected, the better the outcome for the patient,” he explains. “That’s the real advantage of AI: it might draw attention to changes that may otherwise become apparent only in years to come.”
Room for error
However, the system is not foolproof. Dr Schoombee says it may sometimes flag changes that are not significant. When this happens, the radiologist still needs to assess the AI finding, which can take extra time. In general, this is seen as an acceptable trade-off if it helps reduce the risk of missed cancers.
AI also has another limitation: it can’t yet compare current scans with previous ones. This is likely to improve as the technology develops, but for now it remains a challenge. As Dr Schoombee points out, “comparison is a radiologist’s greatest friend”.
What will the future bring?
The key question is: with AI developing at such a rapid pace, what more can we expect?
Big things are ahead, but we must remain realistic, Dr Schoombee cautions. “In South Africa, we don’t use AI tools unless they have recognised approval, such as CE marking, which shows a product meets European safety and health standards, or FDA clearance, as well as any local approvals needed. This helps protect patients and gives doctors confidence in the technology, but it can also delay access to the newest tools.”
Even so, radiology is likely to change significantly in the coming years. AI is getting better at spotting possible problems, deciding which cases need priority, and helping to draft reports.
In future, radiologists may work from AI-generated draft reports, but still review, edit, and approve these reports themselves. This could save time while keeping the radiologist in control.
For now, that is still some way off. “Humans remain very much a part of this process,” Dr Schoombee says. “The reality is that sometimes we agree with AI’s findings, sometimes we don’t. Sometimes they’re important and sometimes they’re not. It’s up to us, as humans, to make the final decision and issue the final report.”