Deep Learning-Based Standard Section Recognition and Multi-Organ Segmentation in Upper Abdominal Ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Upper abdominal ultrasound is crucial for diagnosing various conditions.
• Standardized image acquisition and organ identification are key challenges in ultrasound interpretation.
• Automating these processes can improve efficiency and consistency.

Purpose of the Study:

• To develop and evaluate deep learning models for automated standard section recognition in upper abdominal ultrasound.
• To implement and assess deep learning-based multi-organ segmentation following automated section recognition.
• To investigate the feasibility of a two-stage deep learning framework for upper abdominal ultrasound analysis.

Main Methods:

• A retrospective study utilized 465 upper abdominal ultrasound videos from a single center.
• A two-stage framework involved convolutional neural network-based section recognition followed by deep learning segmentation.
• Models were trained, validated, and tested on 5535 images representing 12 standard upper abdominal sections.

Main Results:

• The standard section recognition model achieved high performance (e.g., 97.00% accuracy).
• Organ segmentation showed strong results for the liver, gallbladder, and right kidney.
• Segmentation performance was lower for the spleen and pancreas, with an IoU of 53.76% for the pancreas.

Conclusions:

• Deep learning models are feasible for standard section recognition and organ segmentation in upper abdominal ultrasound using real-world data.
• The developed models address complementary tasks, potentially aiding in ultrasound interpretation.
• External validation is necessary to confirm the generalizability of these deep learning models.