Deep Active Learning for Automatic Segmentation of Maxillary Sinus Lesions Using a Convolutional Neural Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate segmentation of the maxillary sinus is crucial for dental diagnostics.
• Limited annotated datasets pose a challenge for training deep learning models in medical imaging.

Purpose of the Study:

• To develop and evaluate a deep active learning framework for segmenting maxillary sinus components (bone, air, lesion) in CBCT volumes.
• To assess the accuracy and efficiency of the proposed framework compared to expert segmentation.

Main Methods:

• A three-step active learning framework was implemented using a customized 3D nnU-Net model.
• The framework iteratively improved model performance by incorporating newly annotated volumes.
• Segmentation accuracy was evaluated using Dice Similarity Coefficients (DSCs) for air and lesions.

Main Results:

• The deep active learning framework achieved high DSCs for air segmentation (0.920-0.930) and acceptable DSCs for lesion segmentation (0.750-0.770).
• Significant reductions in segmentation time were observed: approximately 493.2s for 30 scans and 362.7s for 76 scans.
• The framework demonstrated efficient training on limited cone-beam computed tomography (CBCT) datasets.

Conclusions:

• Deep active learning effectively reduces annotation efforts and costs in medical image segmentation.
• The proposed framework offers an efficient solution for training segmentation models with limited CBCT data.
• This approach has the potential to improve the workflow for analyzing dental CBCT scans.