Teacher-student approach for lung tumor segmentation from mixed-supervised datasets.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a leading cause of death, necessitating early detection for improved patient outcomes.
• Automating lung tumor segmentation in radiological images can enhance radiologist efficiency.
• Training deep learning models like convolutional neural networks requires substantial labeled data, which is challenging to obtain in the medical field.

Purpose of the Study:

• To investigate a teacher-student framework for pulmonary tumor segmentation using varied supervision levels.
• To reduce the annotation burden for training automated medical image analysis models.

Main Methods:

• A teacher-student design was employed, where a student model performs end-to-end segmentation and a teacher model provides pseudo-annotated data.
• The framework utilized a combination of limited semantically labeled data and abundant bounding box annotated data.
• The model was trained and evaluated on the MSD Lung dataset for pulmonary tumor segmentation.

Main Results:

• The teacher-student model achieved competitive performance in pulmonary tumor segmentation.
• Performance was comparable between models trained on limited semantic labels and those trained on teacher-annotated data.
• The model trained with minimal semantic labels achieved a mean Dice similarity coefficient of 71.0.

Conclusions:

• Teacher-student designs show promise in reducing annotation requirements for medical image segmentation.
• Less supervised annotation schemes can be implemented without compromising segmentation accuracy.
• This approach can significantly decrease the annotation load in developing automated medical imaging tools.