Diagnostics for pancreatic cancer can improve with the help of AI

Pancreatic cancer is the deadliest form of cancer worldwide, mainly because the disease is often discovered late. Symptoms associated with these tumors are often non-specific, so patients and doctors do not immediately suspect cancer. Moreover, these tumors are difficult to detect in the early stages on abdominal CT scans. Once the diagnosis is made, curative treatment is usually no longer possible. Only 10% of patients survive beyond five years.

Reliable benchmark

AI researcher Henkjan Huisman and radiologist John Hermans explored ways to improve diagnostics using AI. They created a reliable benchmark: a confidential dataset of scans from nearly 400 patients from Western countries, assessed by a large group of international experts. They then invited developers worldwide to submit AI models capable of detecting pancreatic cancer. More than 250 models were submitted, ranked, and compared to experts.

The researchers tested these models on the confidential dataset and found that the best AI models detected pancreatic cancer more accurately than the average radiologist. For example, the AI system produced 38% fewer false positives compared to the group of radiologists. AI made the correct assessment in 92% of scans, versus 88% for the radiologists.

These results show that AI can at least support radiologists in their work and, in the long run, help reduce their workload. The developed AI still needs validation and is not yet available for patients in clinical practice.

But these are important developments, emphasizes lead researcher Henkjan Huisman, “Precisely because we have developed a reliable benchmark, we know that the AI systems outperforming clinicians are truly effective.” The paper is published in The Lancet Oncology journal.

Potential for earlier diagnosis

The best AI model may also offer opportunities for earlier diagnosis. Radiologist John Hermans explains, “In the study, we see initial indications that this AI model could really help achieve faster diagnosis and, therefore, potentially faster treatment. This is a small ray of hope, something we urgently need for this type of cancer.”

However, Hermans stresses that it is still too early to use AI directly for early diagnostics: “We must avoid false positives for this disease, given the unnecessary strain on health care and, especially, the anxiety such suspicions cause for patients.” Therefore, the researchers are now training better AI models on other, broader abdominal scans.