Artificial intelligence used to identify pancreatic cancer in CT scans

In a recent study published in the journal, raysIn Taiwan, researchers have developed a computer-aided detection (CAD) tool based on deep learning (DL) to detect pancreatic cancer in contrast-enhanced abdominal computerized tomography (CT) scans.

Stady: Detecting pancreatic cancer by computed tomography with deep learning: a nationwide population-level study. Image Credit: Suttha Burawonk / Shutterstock

background

Pancreatic cancer patients have the lowest five-year survival rate; Projections indicate that it will emerge as the second leading cause of cancer death in the United States by 2030. In addition, the prognosis for pancreatic cancer rapidly worsens once the tumor grows more than 2 cm, necessitating early detection.

Currently, approximately 40% of tumors less than 2 cm in size are missed by CT pancreatic cancer diagnosis, hampered by the disparity in radiologist expertise. Indeed, there is an urgent, unmet medical need for tools that can enable radiologists to manually analyze pancreatic segmentation to improve the sensitivity of pancreatic cancer detection. Furthermore, in patients with pancreatic cancer, segmentation or identification of the pancreas is difficult because it varies in size and shape and limits many other organs and structures.

In one of their previous work, the researchers showed that a DL-based convolutional neural network (CNN) can accurately distinguish pancreatic cancer from non-cancerous pancreas.

about studying

In this study, researchers tested and validated a similar computer-aided detection (CAD) tool that housed a CNN to segment the pancreas on computed tomography images. In addition, this tool contains a group classifier with five independent classification CNNs to predict the presence of pancreatic cancer. They obtained all analyzed CT scans in the portal venous phase, 70–80 s after intravenous injection of contrast medium.

Training and validation data sets and local and national test data sets were used in the study. The team randomly divided pancreatic cancer patients in a ratio of 8:2 into the training and validation group and the local test group, respectively. They prospectively collected CT studies of 546 patients with pancreatic cancer diagnosed between January 2006 and July 2018 from clinical practices in Taiwan, which constituted their local dataset. These patients were 18 years of age or older with confirmed adenocarcinoma of the pancreas with results registered with the National Cancer Registry. The local data set control group consisted of CT studies of 1,465 individuals with normal pancreas collected between January 2004 and December 2019.

Researchers searched the National Health Insurance (NHI) Principal Disease Certification Registry to retrieve CT scan studies of 669 patients with newly diagnosed pancreatic cancer between January 2018 and July 2019. Similarly, they extracted CT scan studies of 72 kidney and liver donors during the same period. time from the NHI database, which constituted the control group. They also combined these two studies with CT studies of 732 control subjects from the imaging archives of the NHI Database High Referral Center to create a nationwide test dataset for the current study.

Finally, the team trained the five classification CNNs on other subsets of training and validation sets retrieved from the National Institutes of Health Database Tertiary Referral Center, which included CT studies of 437 pancreatic cancer patients and 586 controls. Only when the number of CNNs with a positive prediction equal to or greater than the smallest number yielding a positive probability ratio (LR) greater than one in the validation did the researchers consider that the CT scan showed pancreatic cancer.

The researchers evaluated the performance of the CNN segmentation with a dice score for each patient. Similarly, they evaluated the performance of the CNN classification based on their sensitivity, specificity, and accuracy of each. The team calculated the area under the receiver operating characteristic curve (AUC) and LR. Finally, they used the McNemar test to compare the sensitivities of the CAD instrument and the interpretation of the radiologist.

Results

In the in-house test group, the sensitivity and specificity of the CAD instrument for distinguishing CT malignancies from control studies were 89.7% and 92.8%, respectively, with a sensitivity of approximately 75% for pancreatic cancers smaller than 2 cm. In general, it showed high durability and generalizability. Interestingly, the sensitivity of the CAD instrument was comparable to that of radiologists at a class III academic institution with a large number of pancreatic cancer patients (90.2% vs. 96.1%), suggesting that this instrument may have a higher sensitivity than that of lower radiologists. expertise. It may help reduce the error rate attributable to variances in the radiologist’s experience.

Furthermore, the tool seemed feasible for clinical publication as it provides ample information to assist clinicians. Determine if the images showed pancreatic cancer. She also noted the likely location of the tumor to help radiologists quickly interpret the results. Notably, in about 90% of pancreatic cancers accurately identified by the CAD tool, segmented CNNs correctly locate the tumor. Furthermore, the instrument provided positive CAD LR, a measure of confidence in the classification of pancreatic cancer versus the classification of non-pancreatic cancer to better inform the subsequent diagnostic therapeutic process than a simple binary classification.

Secondary signs in the non-neoplastic portion of the pancreas, including pancreatic duct dilatation, upstream pancreatic parenchymal dystrophy, and abrupt resection of the pancreatic duct, are clues to occult pancreatic carcinomas. A good diagnostic tool should be able to take advantage of these signs in the detection process. In the current study, CNN taxonomic networks correctly classified two cases of pancreatic cancer by analyzing only the non-neoplastic portion of the pancreas by automatically learning secondary markers of pancreatic cancer from the examples.

Conclusions

The new CAD instrument used in the current study demonstrated the potential of a radiologist to complement the early and accurate detection of pancreatic cancers through computed tomography. However, the finding that CNN classification may have learned secondary signs of pancreatic cancer requires further investigation. Similarly, future studies should test the performance of this CAD tool in non-Asian (and Taiwanese) groups to collect data that support its generalizability.

Journal reference:

  • Detecting pancreatic cancer by computed tomography with deep learning: a nationwide population-level study Bo Ting Chen, Tingwei Wu, Bochuan Wang, Dawei Zhang, Cao Lang Liu, Ming Xiang Wu, Holger Roth, Bo Zhang Li, Wei- Chih-Liao, Weichong Wang, Radiology 2022, DOI: https://doi.org/10.1148/radiol.220152And the