Can new deep learning tools enhance CT detection of pancreatic cancer?

Noting that nearly 40% of pancreatic cancer tumors smaller than 2 cm are missed by computed tomography (CT) evaluation, the authors of the new study suggest that new deep learning tools could impact improved detection. suggests that there is a

The study was conducted in Taiwan and published earlier today Radiologyresearchers examined the efficacy of a deep learning tool for detecting malignant pancreatic tumors on contrast-enhanced CT in a nationwide validation test set of 669 participants with pancreatic cancer and 804 participants in a control group. .¹ In the study, a deep learning tool was trained using contrast-enhanced CT scans of 546 pancreatic cancer patients and 733 healthy controls.

Researchers found that deep learning tools had a sensitivity rate of 89.7% and a specificity of 92.8% in detecting pancreatic cancer in a nationwide validation test set. In a local test set data drawn from 109 pancreatic cancer patients and 147 control participants at a tertiary referral center, the study authors found that between radiologist assessment (96.1%) and deep learning tools (90.2%). reported that there was no significant difference in sensitivity.¹

“In this study, we developed an end-to-end deep-learning-based computer-aided detection (CAD) tool that can accurately and reliably detect pancreatic cancer in contrast-enhanced CT scans. ),” study co-author Dr. Wei Chun Wang, of the Center for Artificial Intelligence and Advanced Robotics, National Taiwan University, and colleagues wrote: I’m here.

The results of the study suggest a promising potential for deep learning tools, even for small pancreatic lesions that are often missed by CT, according to the study authors.

“The comparable sensitivity between CAD tools and experienced radiologists in tertiary referral centers supports the idea that CAD tools can help reduce error rates due to expertise mismatch. explained Wang and colleagues. “…Previous studies missed about 40% of <2 cm (pancreatic cancer) on CT scans, whereas our CAD tool has been able to detect <2 cm (pancreatic cancer) in regional and national test sets. cancer) achieved sensitivities of 87.5% and 74.7%, respectively."^1,2

and Accompanying editorialAlex M. Eisen, MD and Pedro Rodrigues, Ph.D., commended “quantity and quality of training and test data”, evaluation of three-dimensional data by algorithms, and robust measures of algorithm effectiveness .³ According to Dr. Aisen and Dr. Rodrigues, emeritus professors of radiology and imaging sciences at Indiana University, using artificial intelligence (AI) to perform routine tasks and track tumor progression over time is “There is potential for widespread adoption in the reading room.” , Clinical Scientist at Philips Healthcare.

Regarding study limitations, the authors acknowledge that sensitivity rate results obtained from tertiary referral centers for prostate cancer may limit extrapolation of the data to a general setting. , also noted the lack of access to radiologist reports from national datasets and the predominantly Asian population of test sets that were fairly homogeneous in terms of ethnic and racial diversity. The control group did not include patients with pancreatic abnormalities other than prostate cancer, according to the study authors.

References

1. Chen PT, Wu T, Wang P, et al. Detection of pancreatic cancer in CT scans by deep learning: a nationwide population-based study. Radiology. available: https://doi.org/10.1148/radiol.220152 Published on September 13, 2022. Accessed on September 13, 2022.

2. DeWitt J, Devereaux BM, Lehman GA, Sherman S, Imperiale TF. A comparison of endoscopic ultrasound and computed tomography for preoperative evaluation of pancreatic cancer: a systematic review. Clin Gastroenterol Hepatol. 2006;4(6):717-25; Quiz 664.

3. Aisen AM, Rodriguez PS. Deep learning for pancreatic cancer detection on CT: artificial intelligence that lives up to the hype. Radiology. available: https://pubs.rsna.org/doi/10.1148/radiol.222126 Published on September 13, 2022. Accessed on September 13, 2022.