Object recognition in medical images via anatomy-guided deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) faces challenges in robust medical image segmentation due to artifacts and pathologies.
• Integrating natural intelligence (NI) with DL offers a promising solution to overcome these limitations.

Purpose of the Study:

• To develop an anatomy-guided deep learning object recognition approach (AAR-DL) that synergistically combines NI and artificial intelligence (AI).
• To achieve expert human-like performance in medical image segmentation, particularly for challenging cases.

Main Methods:

• The AAR-DL approach utilizes four modules, integrating NI at each stage for anatomy modeling and object recognition.
• It employs automatic anatomy recognition (AAR) for initial guidance, followed by DL-based detection using 2D bounding boxes, and refined DL detection guided by an updated anatomy model.
• Spatially sparse and non-sparse objects are handled differently within the DL networks based on anatomy knowledge.

Main Results:

• The AAR-DL approach demonstrated significant error reduction across modules, with location errors of 4.4 mm (thoracic) and 4.3 mm (H&N).
• Performance improvement was particularly dramatic for sparse and artifact-prone objects, where pure DL methods failed.
• AAR-DL achieved accurate recognition, closely matching human performance, highlighting the value of anatomy guidance.

Conclusions:

• High-level anatomy guidance substantially enhances the performance of DL methods in medical image recognition.
• This guidance is especially beneficial for sparse, low-contrast, and artifact-prone objects, improving robustness and stability.
• Anatomy guidance also leads to a significant reduction in DL network training time.