AI detects ‘invisible’ signs of pancreatic cancer years before diagnosis

Researchers believe this technology could help transform pancreatic cancer detection by identifying the disease at a much earlier stage (stage 0), when treatment is more likely to be successful. Instead of being diagnosed once the cancer has progressed to an advanced, potentially terminal stage, patients might have the option to begin treatment while the disease is still treatable.

Although REDMOD outperformed experienced radiologists in the study, more research is needed before it can be introduced into routine medical care. The scientists particularly emphasize the need to test the system in high-risk patients, including those with unexplained weight loss or recently diagnosed diabetes.

Why pancreatic cancer is so difficult to detect early

Pancreatic ductal adenocarcinoma has a low survival rate, largely because it is often not discovered until the disease is advanced. Early pancreatic cancer usually produces few visible symptoms, and the changes it causes in pancreatic tissue may be too subtle to be detected by standard medical imaging. Cancer can also progress quickly, making rapid diagnosis particularly important.

To address these challenges, researchers created an artificial intelligence framework known as Radiomics-Based Early Detection MODel (REDMOD). The system analyzes extremely fine patterns in the texture of pancreatic tissue, known as radiomics, that are not easily seen on conventional computed tomography (CT) scans.

REDMOD also automatically identifies and demarcates the pancreas, separating it from neighboring organs and tissues. This process eliminates the need for specialists to manually trace the boundaries of the organ, reducing inconsistencies that can arise when different people perform the task.

AI discovers hidden cancer signals years before diagnosis

To evaluate the system, researchers examined abdominal CT scans of 219 patients treated at multiple hospitals. At the time of their examinations, the experienced radiologists found no evidence of pancreatic cancer. However, all patients in this group were subsequently diagnosed with the disease.

The scans had been taken at different intervals before diagnosis. For 87 patients (40%), the images were obtained 3 to 12 months earlier. Another 76 patients (35%) had been scanned 12 to 24 months before diagnosis, while 56 (25%) had undergone imaging more than 24 months earlier (up to approximately 3 years). In almost two thirds (64%) of patients, the cancer developed in the head of the pancreas.

The researchers compared these images with CT scans of 1,243 people who remained free of pancreatic cancer up to three years later. The comparison group was matched based on age, gender, and the dates of their scans.

The patients who ultimately developed pancreatic cancer were an average of 69 years old, with ages ranging from 34 to 88 years old. Those in the comparison group were 64 years old on average, with the same age range being 34 to 88 years old.

REDMOD successfully identified subtle imaging signatures associated with pancreatic cancer an average of 475 days before patients received their clinical diagnosis. These warning signs were not previously detected during routine radiological evaluations.

“This time window is of profound significance, as achieving such early detection would significantly increase the likelihood of cure and improve survival,” the researchers point out.

“In fact, modeling studies indicate that increasing the proportion of localized pancreatic ductal carcinomas from 10% to 50% would more than double survival rates, highlighting that the time of diagnosis is the most critical determinant of survival outcomes,” they add.

AI outperforms experienced radiologists

The results also showed that REDMOD was significantly better than experienced radiologists at recognizing early cancer signals.

AI achieved a sensitivity of 73%, compared to 39% for radiologists. Sensitivity measures how effectively a diagnostic method identifies people who actually have the disease, rather than overlooking genuine cases.

The difference was even more pronounced when the researchers looked at exams taken more than two years before the patients’ diagnosis. In these cases, REDMOD achieved a detection rate of 68%, almost three times the 23% recorded by radiologists.

The researchers also assessed how well the system could recognize scans from people without pancreatic cancer. In an independent group of 539 patients from multiple hospitals, REDMOD correctly classified just over 81% of exams as cancer-free. It performed even better on the US National Institutes of Health’s publicly available NIH-PCT dataset (80 patients), correctly identifying 87.5% of them as free of pancreatic cancer.

Additional testing suggested that the warning signs detected by REDMOD were consistent over time. When researchers analyzed scans taken on the same patients several months apart, the system produced matching results 90 to 92 percent of the time. This consistency strengthened the evidence that the AI ​​identified significant changes associated with disease.

Could AI transform early detection of pancreatic cancer?

Despite these promising results, the researchers caution that several questions remain unanswered. One limitation is that the study population did not include a sufficiently diverse range of ethnic backgrounds, which could affect the technology’s performance in different populations.

The system also needs to be prospectively tested, particularly in people at increased risk of pancreatic cancer, before researchers can determine whether it will be effective in everyday clinical practice.

Even with these limitations, the results suggest that AI could potentially help doctors identify pancreatic cancer while it is still at a much more treatable stage, potentially improving survival for patients who would otherwise receive a late diagnosis.

The researchers conclude: “This study validates REDMOD as a fully automated AI framework capable of identifying stage 0 (pancreatic ductal adenocarcinoma) imaging signatures in the normal pancreas, achieving this with substantial turnaround times and superior performance compared to expert radiologists.

They add: “While prospective validation is paramount to confirm clinical utility, the REDMOD framework represents a significant step forward toward paradigm shift for sporadic (pancreatic ductal adenocarcinoma) from late-stage symptomatic diagnosis to proactive preclinical interception, offering tangible hope for improving outcomes in this challenging disease.