Artificial Intelligence Shows Potential for Diagnosing Pancreatic Cancer

Published in JMIR, new research found that the use of artificial intelligence (AI) to diagnose, predict risks of, differentiate, and segment pancreatic cancer led to high levels of accuracy, foreshadowing the role it will continue to play in the future.

According to the National Cancer Institute (NCI), there were about 62,210 estimated new cases of pancreatic cancer in 2022, a condition that has a five-year survival rate of about 11.5 percent.

AI, however, is a resource that continues to grow while making noticeable improvements to the efficiency and quality of healthcare. It often functions by using tools such as electronic health records, computer capabilities, continuous monitoring systems, and large sets of data to transform the diagnostic process.

To gain a better understanding of the capabilities of AI when diagnosing pancreatic cancer, researchers engaged in search queries between June 30 and July 30, 2022, using five online databases. They identified a total of 18,285 articles; however, following a screening process, 30 articles fit in with the inclusion criteria.

All 30 articles were published between 2015 and 2022, 8 of which came from the US and 5 from China. Also, although all related to AI, characteristics varied, as 8 used machine learning, 11 used deep learning, 7 used applied machine learning that included deep learning, and the rest used either adversarial networks or particle swarm intelligence. Regarding models, the articles mainly used convolutional neural networks and logistic regression. Regarding data, 9 used laboratory data, 12 used clinical data, 14 used radiology images, and 1 included demographic data.

Of the 30 articles included in the study, 22 discussed performance measures of the AI models. Of the nine studies that mentioned accuracy, the rate ranged from 71.6 percent to 99 percent, with a mean of 89.4 percent.

Additionally, 12 articles used sensitivity, and 12 articles used specificity as performance measures. The reported sensitivity ranged between 60 percent and 99.9 percent with a mean of 91.3 percent, and the reported specificity ranged between 69.5 percent and 100 percent with a mean of 83.2 percent.

Among the included studies, five separate validation techniques were used, the most common being k-fold cross-validation, used by 10 articles, and external validation used by 10 articles.

Following the collection of results, researchers divided them into four categories. The goal behind this effort was to gain insight from various perspectives.

The first category considered how AI was used for either diagnosis, risk prediction, differentiation, or segmentation. The second category highlighted the differences in AI features, the third focused on the differences in data, and the fourth divided validation methods.

Following the conclusion of this effort, researchers noted that the use of AI to address pancreatic cancer remains understudied. However, they also indicated that this review is a valuable tool that will further assist the scientific community in gaining insight into the use of AI for diagnosing and predicting pancreatic cancer risks.

The exploration of how AI can serve as a tool for treating cancer is not unprecedented.

An example took place in February when researchers from Sylvester Comprehensive Cancer Center at the University of Miami Miller School of Medicine described an AI algorithm that aimed to determine therapeutic targets for glioblastoma multiforme (GBM).

Specifically, researchers used machine learning (ML) to identify two common kinases associated with the tumor progression of two subtypes of GBM.

According to the study, findings such as these indicate how ML algorithms can gain insight into the subjective glioblastoma subtype of a patient and use this data to predict necessary targeted therapies.