An attention base U-net for parotid tumor autosegmentation

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Parotid neoplasms are rare, comprising less than 3% of head and neck cancers.
• Nonspecific imaging features and tumor heterogeneity challenge accurate preoperative diagnosis.
• Automated segmentation tools can potentially improve the evaluation of parotid tumors.

Purpose of the Study:

• To develop and evaluate an attention-based U-net model for automatic parotid tumor segmentation.
• To assess the model's performance against manual segmentation by radiologists.

Main Methods:

• Utilized MRI scans (T1w, T2w, T1wC) from 285 patients with parotid tumors.
• Segmented parotid and tumor tissues manually by three radiologists.
• Trained an attention-based U-net model on a 90% training dataset and validated on a 10% validation dataset using 10-fold cross-validation.

Main Results:

• The model achieved a mean Dice similarity coefficient (DICE) of 0.88 for both parotids.
• Mean DICE for left and right parotid tumors were 0.85 and 0.86, respectively.
• Model performance was comparable to manual segmentation by radiologists.

Conclusions:

• An attention-based U-net model demonstrates high accuracy for parotid tumor autosegmentation on MRI.
• This automated segmentation tool shows potential to assist physicians in evaluating parotid gland tumors.
• The model's performance suggests it can be a valuable aid in the diagnostic workflow for parotid neoplasms.