Patch-wise 3D Segmentation Quality Assessment Combining Reconstruction
and Regression Networks
Abstract
In this article, we propose a novel deep learning framework to identify
location-specific inaccurate segmentation for medical images to aid
downstream analyses. The proposed method relies directly on the
extracted segmentation related features and does not need to use the
independent standard during the inference phase to identify erroneous
regions in the computed segmentation. This is significant because image
segmentation plays an indispensable role for automated quantitative
analysis in modern precision medicine and the absence of case-by-case
ground truth for medical images limits the segmentation quality
assessment. Moreover, in contrast to our approach, to the best of our
knowledge, current deep learning-based segmentation quality assessment
models evaluate the segmentation quality globally and may ignore
significant local segmentation errors in locations where diseases are
prevalent. For the purpose of disease diagnosis and/or prognosis, it is
critical to identify inaccurate segmentation in local regions. We used
two publicly available datasets to evaluate our method and show that our
method is robust and works well to localize small but significant
inaccurate segmentation in the cases of both good and bad overall
segmentation results for different imaging modalities.