In a recent study published in the journal Nature Medicine, a large team of researchers from China, the United States, and the Czech Republic developed a deep learning-based approach to use non-contrast computed tomography (CT) scans for high-accuracy detection and classification of pancreatic lesions for the early detection and treatment of pancreatic ductal adenocarcinoma (PDAC).
Study: Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Image Credit: mi_viri/Shutterstock.com
Background
Pancreatic ductal adenocarcinoma is the most malignant form of solid carcinoma, with a mortality rate of over 450,000 each year. The high mortality rate, however, is largely because PDAC is often detected in the late stages when it is inoperable.
Cases where PDAC is detected incidentally or early have a better prognosis and early treatment often results in substantial improvements in the survival rates of patients.
The median overall survival rate in cases where PDAC has been detected and treated in the early stages is 9.8 years compared to the 1.5-year survival rate for most late-detection cases.
Screening for pancreatic lesions is believed to be the most effective way to detect PDAC in the early stages and significantly lower the mortality rate associated with PDAC. However, given the low prevalence of this form of cancer, mitigation of the over-diagnosis risk requires effective screening techniques with high sensitivity and specificity.
Non-contrast CT has been widely used for the clinical screening of various cancer forms. Combined with artificial intelligence (AI)-based detection and analysis techniques, it can potentially be used for large-scale screening for PDAC.
About the study
In the present study, the team of scientists described an AI-based approach called pancreatic cancer detection with artificial intelligence (PANDA) that can be used to detect and diagnose non-PDAC and PDAC pancreatic lesions accurately using non-contrast CT scans.
This method was developed to use non-contrast CT scans of the chest and abdomen for the detection and diagnosis of PDAC and seven non-PDAC subtypes of lesions, namely, solid pseudopapillary tumor, pancreatic neuroendocrine tumor, mucinous cystic neoplasm, intraductal papillary mucinous neoplasm, chronic pancreatitis, serous cystic neoplasm, and a long list of other non-PDAC pancreatic lesions.
The researchers first internally evaluated the efficiency of PANDA in detecting and diagnosing pancreatic lesions using a set of non-contrast CT scans of the abdomen. PANDA’s performance was compared against that of two reader studies that used non-contrast and contrast CT scans.
In the first study, non-contrast CT pancreatic scans were read by radiology residents, general radiologists, and specialists in pancreatic imaging.
In the second reader study, the performance of PANDA in detecting pancreatic lesions was compared to the performances of specialists in pancreatic imaging who used contrast-enhanced CT scans.
Subsequently, the generalizability of PANDA for various settings was validated using a large multicenter test cohort. Furthermore, chest CT scans were used to test whether PANDA could be used on various patient populations.
The researchers also included chest or abdominal non-contrast CT scans from four settings, namely, outpatient, emergency, physical exam, and inpatient, comprising cumulatively of over 20,500 patients to examine how PANDA could be integrated into large-scale, routine clinical process real-world scenarios.
Results
The results showed that PANDA efficiently detected lesions in the multi-center large-scale validation cohort. Additionally, in specificity and sensitivity, the performance of PANDA was 6.3% and 34.1% greater, respectively, than the average performance of a radiologist in detecting and diagnosing pancreatic lesions.
Furthermore, in the large-scale validation using real-world scenarios for four settings, PANDA achieved 92.9% and 99.9% sensitivity and specificity, respectively.
The researchers demonstrated that a process involving the curation of a large dataset of the common types of pathology-confirmed pancreatic lesions, transfer of lesion annotations from contrast-enhanced CT scans to non-contrast CT images, and use of a deep learning approach to combine diagnostic information modeling for lesions and feedback from real-world scenarios can result in a high-sensitivity and high-specificity detection method for the early diagnosis of pancreatic lesions.
PANDA was also significantly more accurate than radiologists in distinguishing between non-PDAC and PDAC lesions and in differentially diagnosing the eight pancreatic lesion subtypes.
Conclusions
Overall, the findings indicated that PANDA can detect and diagnose non-PDAC and PDAC pancreatic lesions using non-contrast CT scans and distinguish between eight subtypes of pancreatic lesions with high specificity and sensitivity.
These results highlight PANDA’s potential for large-scale screening for pancreatic lesions and the early detection of PDAC.