Early Detection Remains Difficult

Pancreatic cancer is one of the most aggressive and difficult-to-diagnose cancers in medicine. Early detection remains a major challenge because the disease often causes no symptoms in its initial stages and develops deep within the abdomen, making it hard to detect during a routine physical examination.

A Spanish company, Sycai Medical, has taken on the challenge of finding a solution to this problem by creating an AI algorithm capable of analyzing medical images and detecting pancreatic lesions that might go unnoticed during the interpretation of a CT scan. Their approach is based on a simple idea: to leverage medical images that are already routinely obtained to look for abnormalities that may not have been the original reason for the test. In this way, a test performed for any other indication can also become an opportunity to identify a pancreatic lesion that requires follow-up.

From Early Algorithms to a Medical Device

One of the team’s first scientific studies was the development of the algorithm’s first prototype in 2023: an artificial neural network called Attention Gate U-Net designed to detect pancreatic cystic lesions on CT scans. The study was a success as the technology achieved a sensitivity of 93.1% and a specificity of 81.8%. The authors presented it as a tool to support radiologists, not as a substitute for their clinical interpretation.

The technology has continued to evolve toward models capable not only of detecting a lesion but also of providing information about its characteristics, always with the support of healthcare professionals. In 2024, Sycai obtained the European Conformity marking under the European Medical Device Regulation, becoming the first AI solution designed to improve the diagnosis of pancreatic lesions. Currently, the company reports a sensitivity of 96.6% and a specificity of 84.5%.

What Type of AI Does It Use?

The first prototype used artificial neural networks, referring to a branch of machine learning whose operation is based on biological nervous systems — hence the name. It consists of neurons that connect to one another to form several layers:

• Input layer: It receives the initial data and performs initial general processing.
• Hidden layers: These transform the intermediate data through mathematical calculations using weights that indicate the importance of each neural connection.
• Output layer: It delivers the final result. One of the main applications of artificial neural networks is in the field of computer vision: a branch of AI focused on the processing and analysis of images by a computer.

How Does a Neural Network Learn?

The neural network does not initially know what a lesion is. To teach it, just as in human learning, it is provided with many example images in which the location of the lesion is indicated, based on a specialist’s prior analysis. After being fed a large number of images, the network begins to make predictions and compares them with the correct answers.

Based on the accuracy of its predictions, the network itself readjusts the weights associated with its internal connections in order to reduce error. It repeats this process thousands or millions of times with different images. The quality, scope, and diversity of the images used during training are critical. A biased or unrepresentative dataset can cause the algorithm to learn patterns that are not generalizable to other patients or centers. On the other hand, excessive training can lead to overfitting, causing the system to perform very well with known images but lose accuracy when faced with new cases.

The authors of the article disclosed no relevant financial relationships.

This story was translated from Univadis Spain, part of the Medscape Professional Network.