Deep learning-based medical image segmentation with limited labels

Area of Science:

• Radiotherapy
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) auto-segmentation requires extensive labeled data for accurate radiotherapy organ delineation.
• Medical image annotation is labor-intensive, time-consuming, and requires specialized clinical expertise.
• Unlabeled medical images are readily available, presenting an opportunity to leverage this data.

Purpose of the Study:

• To develop a weakly supervised DL approach for medical image segmentation using limited labeled data.
• To leverage abundant unlabeled data by generating pseudo-contours via deformable image registration (DIR).
• To train a robust DL segmentation model that overcomes the limitations of scarce annotated medical images.

Main Methods:

• Proposed a weakly supervised DL training strategy utilizing DIR-based annotations.
• Generated pseudo-contours by propagating atlas contours to unlabeled CT scans using DIR.
• Trained a DL segmentation model using a combination of limited labeled and abundant pseudo-labeled data.

Main Results:

• Achieved high Dice similarity coefficients (e.g., 87.9% for mandible) on the TCIA test set.
• Demonstrated competitive performance on institutional and third-party (PDDCA) datasets.
• Outperformed traditional multi-atlas DIR methods and fully supervised approaches trained with limited data.

Conclusions:

• The proposed weakly supervised DL method effectively utilizes unlabeled data for medical image segmentation.
• This approach shows significant promise for DL-based applications in radiotherapy requiring organ delineation with limited annotated data.
• The DIR-based pseudo-contour generation is a viable strategy to enhance DL model robustness and accuracy.