1232
DISTINGUISHING PANCREATIC CANCER FROM NORMAL PANCREAS AND OTHER PANCREATIC DISEASES ON CT BY ARTIFICIAL INTELLIGENCE: A POPULATION-BASED CASE-CONTROL STUDY
Date
May 21, 2024
Introduction
Approximately 40% of pancreatic cancers (PCs) < 2 cm are missed on CT. Artificial intelligence (AI) has been shown to distinguish PC from normal pancreas and enhance detection of small pancreatic cancers (PCs) on CT. We developed an AI-based computer-aided detection/diagnosis (CAD) tool which could further distinguish PC from other pancreatic diseases on CT and validated the tool in a nationwide population-based case-control study.
Methods
Contrast-enhanced CT of 745 PC patients and 2375 controls, including 706 with other pancreatic diseases from a referral center (National Taiwan University Hospital) were randomly divided (4:1) to train and test the CAD tool, which segmented pancreas and pancreatic lesions by deep learning (DL) and classified whether the pancreas harbored PC by ensembling DL and radiomic analysis with machine learning (ML). The population-based case-control study included newly confirmed PCs throughout Taiwan between January 1, 2018, and December 31, 2020 with eligible CT studies (n=782) and 3,518 controls randomly selected from abdominal CTs obtained during admissions in patients hospitalized throughout Taiwan during the same period (3353 with normal pancreas, 165 with other pancreatic diseases) (Table 1).
Results
In the nationwide population-based study, the CAD tool achieved 91.0% (95% confidence interval: 88.8–93.0) sensitivity and 87.5% (86.4–88.6) specificity [area under receiver operating characteristic curve 0.949 (0.943–0.955)]. Sensitivity for PCs < 2 cm was 82.5% (33/40) in the local test set and 59.7% (37/62) in prediagnostic CTs nationwide including 16 cases that were likely missed/difficult to diagnose. An experienced radiologist rated 100.0% (13/13) and 68.6% (48/70) of the PCs missed by the CAD tool as difficult to diagnose in the local and nationwide test set, respectively. In the local test set, the tool diagnosed 57.1% (4/7) of PCs missed by radiologists, located the tumor in 99.2% (135/136) of true-positives, and informed possibility of false-positive from confounding conditions in 66.7% (28/42) of false-positives.
Conclusion
This CAD tool accurately distinguished PC from normal pancreas and other pancreatic diseases in a nationwide population-based case-control study comprising real-world CT studies. The tool achieved good sensitivity for PCs < 2 cm and informed tumor location and false-positivity. The tool may supplement clinicians in real clinical practice to enhance detection of PC.
Approximately 40% of pancreatic cancers (PCs) < 2 cm are missed on CT. Artificial intelligence (AI) has been shown to distinguish PC from normal pancreas and enhance detection of small pancreatic cancers (PCs) on CT. We developed an AI-based computer-aided detection/diagnosis (CAD) tool which could further distinguish PC from other pancreatic diseases on CT and validated the tool in a nationwide population-based case-control study.
Methods
Contrast-enhanced CT of 745 PC patients and 2375 controls, including 706 with other pancreatic diseases from a referral center (National Taiwan University Hospital) were randomly divided (4:1) to train and test the CAD tool, which segmented pancreas and pancreatic lesions by deep learning (DL) and classified whether the pancreas harbored PC by ensembling DL and radiomic analysis with machine learning (ML). The population-based case-control study included newly confirmed PCs throughout Taiwan between January 1, 2018, and December 31, 2020 with eligible CT studies (n=782) and 3,518 controls randomly selected from abdominal CTs obtained during admissions in patients hospitalized throughout Taiwan during the same period (3353 with normal pancreas, 165 with other pancreatic diseases) (Table 1).
Results
In the nationwide population-based study, the CAD tool achieved 91.0% (95% confidence interval: 88.8–93.0) sensitivity and 87.5% (86.4–88.6) specificity [area under receiver operating characteristic curve 0.949 (0.943–0.955)]. Sensitivity for PCs < 2 cm was 82.5% (33/40) in the local test set and 59.7% (37/62) in prediagnostic CTs nationwide including 16 cases that were likely missed/difficult to diagnose. An experienced radiologist rated 100.0% (13/13) and 68.6% (48/70) of the PCs missed by the CAD tool as difficult to diagnose in the local and nationwide test set, respectively. In the local test set, the tool diagnosed 57.1% (4/7) of PCs missed by radiologists, located the tumor in 99.2% (135/136) of true-positives, and informed possibility of false-positive from confounding conditions in 66.7% (28/42) of false-positives.
Conclusion
This CAD tool accurately distinguished PC from normal pancreas and other pancreatic diseases in a nationwide population-based case-control study comprising real-world CT studies. The tool achieved good sensitivity for PCs < 2 cm and informed tumor location and false-positivity. The tool may supplement clinicians in real clinical practice to enhance detection of PC.
Table 1. Patient characteristics
Presenter
National Taiwan University Hospital
Tracks
Related Products
CLOSTRIDIOIDES DIFFICILE INFECTION INDUCES A PRO-STEATOTIC AND PRO-INFLAMMATORY METABOLIC STATE IN LIVER
BACKGROUND: Recent studies suggest links between _Clostridioides difficile_ infection (CDI) and liver disorders, with non-alcoholic fatty liver disease (NAFLD) increasing CDI risk and CDI exacerbating the progression and prognosis of liver cirrhosis. Moreover, gut dysbiosis, often leading to _C…
