Three-dimensional prostate CT segmentation through fine-tuning of a pre-trained neural network using no reference...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate prostate segmentation on CT scans is crucial for diagnosis and treatment.
• Manual segmentation is labor-intensive and prone to observer variability.
• Deep learning offers potential for automated segmentation but requires large annotated datasets.

Purpose of the Study:

• To develop a computer-assisted method for prostate segmentation on CT images.
• To improve segmentation accuracy and reproducibility using deep learning.
• To overcome limitations of manual annotation by leveraging unannotated data.

Main Methods:

• A base deep learning model was trained on a small, manually segmented CT dataset.
• The model was then fine-tuned using a larger dataset of unannotated CT images.
• Data for pre-training and fine-tuning were acquired from different centers with varying parameters.

Main Results:

• The fine-tuning approach significantly improved validation and testing Dice scores.
• A paired t-test confirmed a statistically significant increase in test scores (p=0.017).
• The method demonstrated effective performance despite variations in CT scanners and imaging parameters.

Conclusions:

• Unannotated CT images can be effectively utilized to enhance automated prostate segmentation models.
• This approach reduces reliance on manual segmentation, saving time and improving consistency.
• The proposed method holds promise for advancing computer-assisted diagnosis and therapy in prostate imaging.