DiSegNet: A deep dilated convolutional encoder-decoder architecture for lymph node segmentation on PET/CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate lymph node (LN) segmentation is crucial for cancer diagnosis from medical images.
• Challenges include low contrast and variations in LN size and shape.
• Automated methods are needed to improve diagnostic assessment.

Purpose of the Study:

• To develop a novel deep learning method for automated lymph node segmentation in PET/CT images.
• To address challenges of low contrast and anatomical variability in LN segmentation.
• To improve the accuracy of lymph node segmentation for cancer assessment.

Main Methods:

• A new neural network architecture, DiSegNet (Dilated SegNet), was designed.
• Incorporated a cosine-sine (CS) loss function to handle voxel class imbalance.
• Utilized a multi-stage, multi-scale Atrous spatial pyramid pooling (MS-ASPP) sub-module for multi-scale information processing.

Main Results:

• DiSegNet achieved an average Dice similarity coefficient of 77% using the CS loss function.
• This performance surpasses the baseline SegNet method (71% Dice similarity coefficient with cross-entropy loss).
• Evaluated using four-fold cross-validation on 63 PET/CT datasets.

Conclusions:

• The proposed DiSegNet with CS loss demonstrates superior performance for lymph node segmentation.
• This method shows significant potential for clinical applications in disease quantification.
• Automated segmentation can aid in more accurate cancer diagnosis and treatment monitoring.