Automated classification of cervical lymph-node-level from ultrasound using Depthwise Separable Convolutional Swin...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical lymph node classification is crucial for disease diagnosis and treatment but currently relies on subjective physician experience, leading to misclassification.
• Existing ultrasound methods struggle with subtle inter-class differences in lymph node levels (II-IV) due to anatomical variations like sternocleidomastoid muscles.

Purpose of the Study:

• To develop an automated, accurate cervical lymph node-level classification system using deep learning.
• To address challenges in ultrasound image quality and data imbalance for improved diagnostic performance.

Main Methods:

• Proposed a novel Depthwise Separable Convolutional Swin Transformer model integrating deepwise separable convolution into self-attention for discriminative local feature capture.
• Developed a new loss function to mitigate data imbalance issues.
• Designed a unified pre-processing algorithm to handle variations in ultrasound image quality (low contrast, blurring) from different devices.

Main Results:

• The model achieved an average accuracy of 80.65% and an F1 score of 79.42% on a dataset of 1146 cervical ultrasound lymph node cases after five-fold cross-validation.
• Validation metrics included precision (80.68%), sensitivity (78.73%), and specificity (95.99%).
• Visualization confirmed the model's focus on expert-identified regions of interest (ROIs).

Conclusions:

• The proposed Depthwise Separable Convolutional Swin Transformer demonstrates significant potential for accurate automated cervical lymph node-level classification.
• The model's ability to identify key local features and handle data imbalance offers a promising advancement over subjective diagnostic methods.
• The approach shows clinical relevance by aligning with expert-defined regions of interest in ultrasound imaging.