Pancreatic cancer is a high-grade malignant intracranial tumor that arises from the uncontrolled proliferation of pancreas cells. In the early stages, patients with pancreatic cancer do not feel ill. Once they begin to feel ill, the disease often progresses to the middle and late stages.
Typically, the clinical diagnostic methods of pancreatic cancer can be divided into the following categories:
Despite the availability of the above diagnosis methods, how to achieve an early diagnosis of cancer is still a great challenge.
As early pancreatic tumors are small and hidden, they can hardly be detected even on CT scans. In the paper Goenka and colleagues recently published in Gastroenterology, they creatively introduced artificial intelligence (AI) into the pre-diagnostic detection of pancreatic cancer. They built four different radiomics-based machine-learning models for the differentiation of healthy people and pancreatic cancer patients.
After that, they further assessed the specificity of the model with the highest accuracy among the four models by using the information from two different Health datasets. Meanwhile, they also set up a control group by choosing two radiologists to evaluate the situation of the pancreas independently.
Fig.1 The schematic diagram of the testing process. (Mukherjee S, 2022)
As shown in Figure 1, they recruited 155 patients who happened to have had CT scans in the past three years before the diagnosis and 265 age-matched healthy bodies to have this CT scan analysis. The result confirmed that the support vector machine model was the best-performing model (95.5% sensitivity and 90.3% specificity). Those models can accurately predict (accuracies:94~98%) future risk of pancreatic cancer at a median time of 386 days before clinical diagnosis. Meanwhile, these results were unaffected by diverse variations such as image noise, scanner models and processing parameters. Moreover, results from model groups and the radiologist diagnosis group were also evaluated for accuracy. There were false positive findings in the radiologist's diagnosis, compared with none in the model group. The reason for this consequence is that computers can identify imaging features that the human eye cannot detect.
Although the effectiveness of this method has been proved, there are still great challenges to translating this method into clinical practice. The precise application of AI needs to rely on a large number of databases, but the amount of data available is far from enough. Goenka's team keeps exploring the option of further validating the AI models on CTs for the early diagnosis of pancreatic cancer and may be able to achieve an ultra-early clinical diagnosis of pancreatic cancer soon.
Reference
For Research Use Only.