Head and Neck Cancer Segmentation in FDG PET Images: Performance Comparison of Convolutional Neural Networks and...

Area of Science:

• Medical image analysis
• Machine learning in oncology
• Radiomics and computational pathology

Background:

• Convolutional neural networks (CNNs) excel in medical image segmentation.
• Vision Transformers are emerging for computer vision tasks, including segmentation.
• Inductive bias in machine learning models is critical for medical image segmentation with limited data.

Purpose of the Study:

• To quantitatively compare the performance of CNN-based and Transformer-based networks for head and neck cancer (HNC) lesion segmentation in [F-18] fluorodeoxyglucose (FDG) PET scans.
• To assess the impact of inductive bias on segmentation performance in a clinical trial dataset.
• To evaluate the effectiveness of attention mechanisms in CNNs for HNC lesion segmentation.

Main Methods:

• Performance evaluation of U-Net, U-Net-CBAM (CNNs), UNETR, TransBTS, and VT-UNet (Transformers).
• Utilized 272 diverse FDG PET-CT scans from the ACRIN 6685 clinical trial, encompassing 650 lesions (primary and secondary).
• Employed multiple error metrics for quantitative performance analysis.

Main Results:

• CNN-based approaches achieved superior performance, with Dice coefficients ranging from 0.833 to 0.809.
• U-Net-CBAM showed improved segmentation for smaller lesions compared to the standard U-Net, highlighting the benefit of attention mechanisms.
• The study identified image features relevant for HNC lesion segmentation.

Conclusions:

• CNNs, particularly U-Net variants with attention, remain highly effective for HNC lesion segmentation in FDG PET scans.
• Transformer-based models show potential but require further optimization for this specific application.
• Accurate segmentation performance estimation necessitates the inclusion of both primary and secondary lesions to avoid bias.