Enhanced IDOL segmentation framework using personalized hyperspace learning IDOL | JoVE Visualize

Area of Science:

Medical Imaging
Radiotherapy
Artificial Intelligence

Background:

Adaptive radiotherapy (ART) aims for precise dose delivery and tissue sparing but is hindered by time-consuming manual recontouring.
Deep learning-based segmentation (DLS) shows promise for automating contouring but requires large, high-quality datasets for generalizability.
Existing DLS methods struggle with clinical implementation due to challenges in data curation and achieving patient-specific accuracy.

Purpose of the Study:

To introduce the personalized hyperspace learning (PHL)-IDOL framework to enhance auto-segmentation for ART workflows.
To address limitations of previous intentional deep overfit learning (IDOL) by enabling overfitting to specific patient characteristics.
To generate patient-specific datasets that improve model performance in medical image segmentation.

Main Methods:

A two-stage training process: first, a general DLS model is trained on diverse patient data (n=100).
The general model is then fine-tuned using a personalized dataset created by selecting similar patients (based on MSE, PSNR, SSIM, UQI) and deforming their contours.
Performance was evaluated by comparing Dice Similarity Coefficient (DSC) and Hausdorff distance 95% (HD95%) against general, continual, and conventional IDOL models using 18 structures in 20 test patients.

Main Results:

The PHL-IDOL framework significantly improved segmentation performance, achieving an average Dice score of 0.87 (compared to 0.81-0.83 for other models).
Hausdorff distance 95% was reduced to 2.36 with PHL-IDOL, outperforming other methods (3.06-2.79).
Standard deviations in performance metrics were nearly halved with PHL-IDOL compared to the general model, indicating increased consistency.

Conclusions:

The PHL-IDOL framework demonstrates superior auto-segmentation performance compared to general DLS approaches.
Leveraging patient-specific information via PHL-IDOL is crucial for advancing online adaptive radiotherapy workflows.
This approach holds significant promise for improving the efficiency and accuracy of ART.