Generalizability, robustness, and correction bias of segmentations of thoracic organs at risk in CT images.

Abstract

Objective: This study aims to assess and compare two state-of-the-art deep learning approaches for segmenting four thoracic organs at risk (OAR)-the esophagus, trachea, heart, and aorta-in CT images in the context of radiotherapy planning.

Materials and methods: We compare a multi-organ segmentation approach and the fusion of multiple single-organ models, each dedicated to one OAR. All were trained using nnU-Net with the default parameters and the full-resolution configuration. We evaluate their robustness with adversarial perturbations, and their generalizability on external datasets, and explore potential biases introduced by expert corrections compared to fully manual delineations.

Results: The two approaches show excellent performance with an average Dice score of 0.928 for the multi-class setting and 0.930 when fusing the four single-organ models. The evaluation of external datasets and common procedural adversarial noise demonstrates the good generalizability of these models. In addition, expert corrections of both models show significant bias to the original automated segmentation. The average Dice score between the two corrections is 0.93, ranging from 0.88 for the trachea to 0.98 for the heart.

Conclusion: Both approaches demonstrate excellent performance and generalizability in segmenting four thoracic OARs, potentially improving efficiency in radiotherapy planning. However, the multi-organ setting proves advantageous for its efficiency, requiring less training time and fewer resources, making it a preferable choice for this task. Moreover, corrections of AI segmentation by clinicians may lead to biases in the results of AI approaches. A test set, manually annotated, should be used to assess the performance of such methods.

Key points: Question While manual delineation of thoracic organs at risk is labor-intensive, prone to errors, and time-consuming, evaluation of AI models performing this task lacks robustness. Findings The deep-learning model using the nnU-Net framework showed excellent performance, generalizability, and robustness in segmenting thoracic organs in CT, enhancing radiotherapy planning efficiency. Clinical relevance Automatic segmentation of thoracic organs at risk can save clinicians time without compromising the quality of the delineations, and extensive evaluation across diverse settings demonstrates the potential of integrating such models into clinical practice.