LDiT3D: Head and Neck Tumor Segmentation Using a 3D Latent Diffusion Transformer Model

Head and neck (H&N) tumors are among the most common cancers worldwide. However, manual annotation of H&N tumors for each patient poses a significant burden on clinicians. Therefore, developing automated and accurate methods for H&N tumor segmentation is of great clinical importance.

We introduce LDiT3D, a novel 3D transformer-based diffusion model designed to segment H&N tumors from 3D Fluorodeoxyglucose positron emission tomography (FDG-PET) and X-ray computed tomography (CT) scans, primarily focusing on oropharyngeal tumors. LDiT3D utilizes stacked 3D vision transformer blocks to effectively capture spatial context in volumetric data. To address the high computational cost typically associated with 3D transformers, we incorporate a pre-trained autoencoder that projects the input into a lower-dimensional latent space where the diffusion process is performed more efficiently. In addition to the diffusion loss computed in latent space, a task-specific segmentation loss is introduced in pixel space to further enhance segmentation accuracy.

Methods based on U-Net and transformer architectures were selected as reference methods. Comparative results demonstrated that the proposed LDiT3D achieved highly accurate segmentation performance, with the highest Dice Score and lowest Hausdorff distance on both public and local datasets. Moreover, LDiT3D showed a more balanced performance between Precision and Recall, as indicated by the highest Fowlkes-Mallows Score. Ablation studies further confirmed the effectiveness of the key components in the LDiT3D design.

LDiT3D delivers more accurate and balanced H&N tumor segmentation compared to reference methods.

These findings highlight LDiT3D's potential to improve automated H&N tumor annotation and advance medical image analysis.