Open-source domain adaptation to handle data shift for volumetric segmentation-use case kidney segmentation

Current development of kidney segmentation models has focused on using a single-phase CT, resulting in significant performance degradation caused by simple characteristic drift in testing datasets, e.g., difference in contrast phase appearance.

We introduce a domain adaptation approach leveraging a latent space discriminator to train a robust model for segmenting kidneys from CT volume irrespective of the contrast dose and functional anomaly. We aim to handle three primary domain shifts between training and testing datasets-(i) contrast to non-contrast, (ii) arterial to venous phase, and (iii) normal to abnormal kidney.

Our model is trained on two publicly available non-contrast and arterial phase image datasets and validated on both public (KiTS21 and STU) and private (Mayo Clinic) datasets with distinct contrast phases and abnormality in the kidney. On all four datasets with domain shift, the proposed model achieved a 0.8892 DICE score, and interestingly, it outperformed the baseline models, including TotalSegmentator, and popular domain adaptation methodologies on the external validation.

Evaluation of internal and external tests demonstrates improved segmentation quality with domain adaptation while leveraging less data than the baseline. An open-source codebase can be accessed.

Question Variations in contrast phase uptake are challenging for evaluating impaired kidney function due to differences in imaging appearance. Findings The proposed open-source domain adaptation approach for kidney segmentation from CT volumes, handles domain shifts to accurately measure kidney volume regardless of contrast dose or functional anomaly. Clinical relevance The domain-shift resilient kidney segmentation of the volumetric CT images is crucial for patients as it ensures accurate and automated assessment of kidney health, irrespective of contrast uptake, enabling timely diagnosis and personalized treatment plans, regardless of contrast uptake.