Author: John Ginn, Chenlu Qin, Deshan Yang π¨βπ¬
Affiliation: Duke University, Department of Radiation Oncology, Duke University π
Purpose: Clinical implementation of auto-segmentation tools has been hindered by poor interpretability and generalizability of AI models, necessitating the development of automated contour quality assurance (QA) systems. We propose a novel contour QA workflow that involves sampling small image patches from an unannotated scan, retrieving the most similar patches from a database of annotated scans, and comparing the auto-contours in the former patches with the manual contours in the latter. To lay the foundation, this study aims to quantify the visual similarity between 3D CT image patches based on radiomic features.
Methods: Fifty abdominal CT scans from the publicly available TotalSegmentator dataset were studied. From each scan, 900 pairs of small patches were sampled along the borders of organs. For each patch, two types of radiomic features were examined separately: 185 handcrafted features available through the MATLAB Medical Imaging Toolbox, and 64 deep features extracted using a convolutional neural network. Regression models were trained to predict the distance between the center of two patches from the absolute differences between the extracted feature values. The performance was evaluated by mean squared error (MSE), squared correlation coefficient (r2), and Spearmanβs rank (Ο).
Results: With handcrafted radiomic features, the XGBoost regressor achieved MSE = 3.537, r2 = 0.777, and Ο = 0.850; with deep radiomic features, the Siamese network achieved MSE = 9.117, r2 = 0.437, and Ο = 0.687. Additionally, to improve storage efficiency, feature selection methods were implemented; to improve model accuracy and robustness, predictions were rejected if the predicted distances were inconsistent when both patches were shifted by 1β2 voxels in the same direction.
Conclusion: The study demonstrates the viability of quantifying patch similarity (approximated as the distance between patch centers) using handcrafted and deep radiomic features. It will enable content-based image patch retrieval for the proposed contour QA workflow.