Attention Guided Lymph Node Malignancy Prediction in Head and Neck Cancer

Area of Science:

• Medical Imaging
• Oncology
• Machine Learning

Background:

• Accurate lymph node (LN) malignancy classification is crucial for head and neck cancer (HNC) treatment planning.
• Deep learning models often require large datasets, which are limited in real-world medical applications.

Purpose of the Study:

• To develop an attention-guided classification (AGC) scheme for accurate LN malignancy status prediction in HNC.
• To incorporate human knowledge (LN contours) into model training to reduce reliance on large sample sizes.
• To enable malignancy determination without requiring precise LN delineation during inference.

Main Methods:

• An attention-guided convolutional neural network (agCNN) module identifies discriminative regions within a region of interest (ROI).
• A classification convolutional neural network (cCNN) module predicts LN malignancy probability using the highlighted features.
• The scheme was validated on PET and contrast-enhanced CT data from 129 HNC patients (791 LNs) using 5-fold cross-validation.

Main Results:

• The AGC scheme achieved high performance metrics: 0.91 sensitivity, 0.93 specificity, 0.92 accuracy, and 0.98 area under the ROC curve.
• Significantly outperformed conventional convolutional neural network and radiomics approaches.
• Effectiveness demonstrated in highlighting discriminative regions for malignancy prediction.

Conclusions:

• The developed AGC scheme effectively predicts LN malignancy in HNC.
• It outperforms traditional radiomics and standard CNN models.
• The method's ability to highlight discriminative regions is key for accurate malignancy determination.