Deep Learning for Automatic Segmentation of Pituitary Adenomas: A Videomics Study.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Surgery
• Neurosurgery

Background:

• Videomics integrates video-endoscopy and AI for real-time surgical analysis.
• Accurate intraoperative segmentation of pituitary adenomas (PAs) is vital for precision in endoscopic surgery.
• Deep learning (DL) models offer potential for automated PA segmentation.

Purpose of the Study:

• To evaluate and compare the performance of Swin Transformer, YOLO, and Mask R-CNN DL models for automated PA segmentation.
• To determine the most accurate model for delineating PA tumor boundaries in endoscopic surgery.
• To assess the potential of AI in enhancing videomics for neurosurgical procedures.

Main Methods:

• Retrospective analysis of 700 frames from endoscopic endonasal surgeries for PAs (Jan 2022-Dec 2023).
• Manual segmentation by two clinicians (κ = 0.85) for ground truth.
• Training and optimization of Swin Transformer, YOLOv8x, and Mask R-CNN models over 100 epochs using mean Average Precision (mAP).

Main Results:

• Swin Transformer achieved the highest mAP[0.50] (0.607), significantly outperforming YOLOv8x (0.416) and Mask R-CNN (0.480) (P < .05).
• Swin Transformer demonstrated superior Dice Similarity Coefficient (0.89), recall (0.91), and precision (0.88) compared to other models.
• Statistical analysis confirmed significant performance differences between the models (P < .05).

Conclusions:

• The Swin Transformer model shows superior accuracy for pituitary adenoma boundary delineation.
• This DL model holds significant potential as an advanced tool for intraoperative PA segmentation.
• AI-driven videomics can enhance precision in endoscopic endonasal surgeries.