Multi-stream and multi-scale fusion rib fracture segmentation network based on UXNet

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of rib fractures is crucial for surgical interventions and patient recovery.
• Segmenting rib fractures in CT images is challenging due to their irregular morphology and variable positioning.
• Existing methods struggle with the precise identification of elongated rib fractures.

Purpose of the Study:

• To enhance the accuracy of elongated rib fracture segmentation in computed tomography (CT) images.
• To improve the efficiency of clinical diagnosis for rib fractures.
• To develop a deep learning model that overcomes segmentation challenges posed by irregular fracture morphology.

Main Methods:

• Proposed a multi-stream and multi-scale fusion network based on efficient attention UXNet (M2SUXNet).
• Introduced the multi-stream and multi-scale fusion (M2SF) module for feature extraction, integrating multi-scale and channel information.
• Employed an efficient attention (EA) module to combine channel and spatial features, enhancing focus on fractures.
• Utilized a joint loss function (BCE with Logits Loss and Dice Loss) to address sample imbalance.

Main Results:

• The M2SUXNet model achieved a Dice coefficient of 75.34% on the RibFrac dataset.
• The model obtained a joint intersection over union (IoU) of 60.44%.
• A precision of 93.79% was recorded, demonstrating high accuracy in fracture identification.

Conclusions:

• The proposed M2SUXNet model demonstrates superior accuracy and feasibility compared to existing methods for rib fracture segmentation.
• M2SUXNet effectively improves segmentation performance for elongated rib fractures.
• The model contributes to more efficient and accurate clinical diagnosis of rib fractures.