A VVBP data-based pancreatic lesion detection model with noncontrast CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Pancreatic cancer (PC) is highly aggressive, necessitating improved early detection methods.
• Noncontrast CT (NCCT) imaging provides a foundation for developing early diagnostic algorithms.
• View-by-view back-projection (VVBP) data from CT scans contain rich, often overlooked, complementary information.

Purpose of the Study:

• To propose and evaluate a novel AI model for pancreatic lesion detection using NCCT-based VVBP data.
• To enhance early diagnosis, prognosis, and survival rates for pancreatic cancer.
• To investigate the effectiveness of integrating VVBP data for improved feature representation.

Main Methods:

• A deep learning model integrating ResNet50-Unet, a multicross channel-spatial-attention (mcCSA) mechanism, and Faster R-CNN with weighted candidate bounding box fusion (WCBF).
• Processing VVBP data into N sparse images (optimal N=3) for feature extraction and fusion.
• Utilizing a reconstruction branch to compensate for information loss in simulated VVBP data.

Main Results:

• The proposed model achieved high performance metrics, including 90.5% patient-level recall and 76.9% AP50.
• The model outperformed competing methods in detecting pancreatic lesions.
• Optimal performance was observed when processing VVBP data into 3 sparse images.

Conclusions:

• The developed AI tool demonstrates significant clinical potential for early pancreatic cancer detection and screening.
• Leveraging VVBP data offers a richer representation for improved lesion detection accuracy.
• The model's high patient-level recall suggests its utility in real-world clinical settings.