Multiscale and multiperception feature learning for pancreatic lesion detection based on noncontrast CT.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Pancreatic cancer has a high mortality rate due to challenges in early diagnosis and treatment.
• Current diagnostic methods struggle with low contrast and complex anatomy in noncontrast CT (NCCT) images.

Purpose of the Study:

• To create an intelligent, noncontrast CT (NCCT)-based model for early pancreatic cancer detection.
• To overcome limitations of NCCT imaging, such as low contrast and complex anatomical structures.

Main Methods:

• Developed a multiscale and multiperception (MSMP) feature learning network using ResNet50 and a feature pyramid network.
• Incorporated multiscale atrous convolutions, contextual attention, and channel/spatial attention for enhanced feature extraction.
• Utilized Faster R-CNN for lesion detection on NCCT image patches of the pancreas.

Main Results:

• The MSMP network significantly improved detection performance over conventional methods.
• Achieved recall of 90.95%, precision of 68.21%, F1 score of 77.96%, F2 score of 85.26%, and Ap50 of 70.14% at the patient level.
• Demonstrated effective mining of NCCT imaging features for pancreatic lesion detection.

Conclusions:

• The proposed MSMP-based NCCT detection model shows strong potential for early pancreatic cancer diagnosis.
• This intelligent tool could enhance early detection rates and improve patient prognosis.
• Further development could lead to a widely applicable clinical tool for pancreatic cancer screening.