MDU-Net: A Convolutional Network for Clavicle and Rib Segmentation from a Chest Radiograph.

Area of Science:

• Medical Image Analysis
• Deep Learning in Radiology
• Computer-Aided Diagnosis

Background:

• Automatic bone segmentation in chest radiographs is crucial but challenging due to artifacts and tissue shadows.
• Traditional methods struggle with accuracy, and a lack of annotated data hinders deep learning approaches for clavicle and rib segmentation.
• Existing deep learning models have shown success in segmenting other organs but not extensively for chest bones.

Purpose of the Study:

• To develop an accurate deep learning model for segmenting clavicles and ribs from chest X-rays.
• To address the challenge of insufficient annotated datasets for bone segmentation in chest radiography.
• To improve the accuracy and reliability of automatic bone segmentation in medical imaging.

Main Methods:

• A novel multitask dense connection U-Net (MDU-Net) was developed, combining U-Net's feature fusion, DenseNet's connectivity, and a multitasking mechanism.
• A mask encoding mechanism was introduced to enhance the learning of background features.
• Transfer learning was employed to improve feature extraction capabilities.
• A new dataset of chest X-rays with detailed bone annotations was created.

Main Results:

• The MDU-Net achieved high average Dice Similarity Coefficients (DSC): 93.78% for clavicle, 80.95% for anterior ribs, 89.06% for posterior ribs, and 88.38% for all bones.
• The model demonstrated robust performance across fourfold cross-validation on 88 chest radiography images.
• The proposed mask encoding and transfer learning strategies contributed to improved segmentation accuracy.

Conclusions:

• The MDU-Net model offers a significant advancement in automatic bone segmentation from chest radiographs.
• The developed dataset and MDU-Net architecture effectively overcome previous limitations in accuracy and data availability.
• This approach holds promise for enhancing diagnostic capabilities in radiology through improved medical image analysis.