Patient specific deep learning based segmentation for magnetic resonance guided prostate radiotherapy.

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Daily contouring is essential for MR-Linac radiotherapy treatments.
• Deformable image registration (DIR) struggles with significant anatomical changes during treatment.
• Artificial neural networks (ANNs) require large datasets, posing a challenge for contouring.

Purpose of the Study:

• To propose patient-specific ANNs for contouring in adaptive MR-Linac workflows.
• To address the challenge of limited training data in ANN-based contouring.
• To evaluate the performance of patient-specific ANNs against DIR methods.

Main Methods:

• Developed U-net shaped ANN models trained on initial treatment fraction images for each patient.
• Applied patient-specific ANNs to subsequent treatment fraction images for contouring.
• Compared ANN contouring accuracy (Dice coefficient, Added Path Length) with manual contours and DIR algorithm results.

Main Results:

• ANN models achieved Dice coefficients of 0.92 ± 0.03 (CTV), 0.93 ± 0.07 (bladder), and 0.84 ± 0.10 (rectum).
• DIR achieved Dice coefficients of 0.95 ± 0.03 (CTV), 0.93 ± 0.08 (bladder), and 0.88 ± 0.06 (rectum).
• Added Path Length (APL) values were comparable between ANN and DIR methods for all structures.

Conclusions:

• Patient-specific ANNs demonstrate comparable accuracy to clinical DIR for contouring in MR-Linac treatments.
• This approach effectively mitigates data scarcity issues for ANN training.
• Patient-specific ANNs show promise for adaptive radiotherapy workflows.