ESTRO 2024 - Abstract Book

S3140

Physics - Autosegmentation

ESTRO 2024

Fig. 2. Framework of the deep learning-based auto-segmentation with the proposed loss scheme.

Results:

Table 1 illustrates the performance of the proposed region-specific loss and adaptive region-specific loss in terms of the Dice similarity coefficient (DSC), recall, precision, and average surface distance (ASD) on public and in-house head and neck datasets, outperforming a conventional baseline loss (Dice loss + cross-entropy loss). Moreover, the DSC results of more than 90% of the organs across different body sites are greatly improved. Qualitative analyses of the network prediction also demonstrate the effectiveness of the proposed loss scheme in clinical auto-segmentation.

Table 1. Organ-averaged evaluation results on the public and in-house datasets by using different loss functions.

Public Dataset

In-house Dataset

ASD

ASD

Loss

DSC

Recall

Precision

DSC

Recall

Precision

(mm)

(mm)

Baseline

0.751

0.763

0.770

1.087

0.657

0.729

0.632

12.786

Region Specific

0.764

0.778

0.778

1.032

0.669

0.754

0.636

12.604

Adaptive Region Specific

0.772

0.801

0.766

1.026

0.685

0.760

0.652

11.283

Conclusion:

This study introduces a novel concept of local loss and proposes an adaptive region-specific loss for more effective auto-segmentation network learning, without modifying the network architecture or requiring additional data preparation. The technique is very versatile and can be extended to various tasks, such as image auto-registration, promising a generalized solution in medical image processing and fostering improved deep learning decision-making in radiotherapy.