ESTRO 2024 - Abstract Book

S3139

Physics - Autosegmentation

ESTRO 2024

Stanford University, Radiation Oncology, Stanford, USA

Purpose/Objective:

The deep learning method has achieved great success in medical image auto-segmentation. However, the technique often falls short in certain areas that are challenging to delineate, such as those near an organ boundary with low tissue contrast, which significantly limits the performance and clinical workflow of current deep learning networks. This study explores a general network optimization approach and proposes a novel loss scheme to substantially improve the performance of existing auto-segmentation techniques.

Material/Methods:

In deep learning, a loss function is used to measure the discrepancy between the network prediction and the ground truth to guide network optimization, which critically determines the success of deep learning modeling. Different from conventional loss functions, which weigh all regions in the medical image volume equally, we first propose a region specific loss to implicitly place more importance on difficult-to-predict sub-regions based on the network prediction. Specifically, we divide the entire image volume into multiple sub-regions, each with an individualized loss constructed to achieve region-specific enhancement for different sub-regions for optimal local performance, as shown in Fig.1. Furthermore, an adaptive region-specific loss is developed by dynamically adjusting the optimization emphasis of the region-specific loss between false positive and false negative prediction errors in each sub-region during the training process to enhance the overall prediction accuracy. Both public and in-house CT datasets of over 200 cases of different body sites are used to evaluate the proposed loss scheme in multi-organ segmentation.

Fig. 1. Illustration of the conventional Dice loss (left) and the proposed loss designs, including the region-specific loss (middle) and adaptive region-specific loss (right).