V Leandro Alves¹*, M H Soomro¹, J Siebers¹, (1) University of Virginia Health System, Charlottesville, VA

• 
  V Leandro Alves

SU-IePD-TRACK 2-4 (Sunday, 7/25/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: To train a classification model that provides a class-probability estimate for labeling delineated Head and Neck (HN) organs at risk (OARs).

Methods: A knowledge-based quality control system composed of an extremely-randomized-trees classifier (ERT) model trained using segmentations from 234 publicly available CT Head and Neck image-sets were used to solve an Organ-at-Risk (OAR) classification problem. The feature vector for each OAR comprises shape, image intensity, location, and orientation features. The dataset was augmented using SMOTE. 5-fold stratified cross-validation was used to obtain learning curves for ERT k-nearest neighbors vote (KNN) and Dummy Classifiers (DC) for a random classification baseline. Each model was evaluated using a stratified 5-fold iterator with different randomization in three repetitions. ERT output was further derived from the probability of an OAR being an HN organ. A probability threshold was selected that maximizes the detection recall, consequently minimizing the fraction of false negatives. The dataset contains 28 OAR delineations. The model was optimized to classify 12 common HN structures, and the remaining ones were kept as hold-out OARs.

Results: The cross-validated recall was DC 0.09±0.02, KNN 0.98±0.01, and ERT 0.994±0.005. The training curves for ERT showed convergence after a sample size of around 500 OARs, showing excellent recall scores in both train and validation samples as a low probability of overfitting. The mean classification probability for in-sample and hold-out OARS were respectively 0.97±0.07 and 0.41±0.08. A minimum classification probability of 0.77 was obtained using grid-search, resulting in a global 0.96 recall score

Conclusion: ERT classifies HN OARs with high recall and provides superior results compared to KNN and DC. ERT reduces the likelihood of ROI delineation mislabeling and might detect aberrant delineations. It enables mass ROI name standardization with low miss-classification errors for data-mining of heterogeneous historical priors.

Funding Support, Disclosures, and Conflict of Interest: NIH R01CA222216

ePosters

Keywords

Machine Learning

Taxonomy

TH- Dataset Analysis/Biomathematics: Machine learning techniques

Contact Email