(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purposes: The advancement of highly conformal and targeted radiotherapy delivery mechanisms imposes an urgent need for automatic segmentation with high accuracy and efficiency, where deep networks (DNN) had yielded great promise. In addition, it is important to understand the confidence associated with such contour set. While Monte Carlo Dropout (MCDO) approaches with Bayesian network setting can yield uncertainty estimation, we hypothesize that treating the contours in a pseudo-probabilistic setting could yield comparable information with much better computational effectiveness.
Methods: We utilize the soft membership assignment from intermediate layers of segmentation DNNs as "free" by product, and endow it with statistical interpretations to define uncertainty. Specifically, trace of estimated multivariate covariance and multiple variant entropy were investigated as uncertainty indicators. 48 head and neck volumes were randomly split into the Training (27 volumes, 1295 slices), validation (9 volumes, 476 slices) and testing (12 volumes, 602 slices). A 2D denseUnet was used as the common basic segmentation network structure.
Results: The benchmark MCDO SegNet and the two proposed uncertainty indicators all corroborate with the error in labeling, with AUC >0.9. However, the required MCDO process in segnet induces both performance dependence on the number of process trajectories as well as the time cost. While a Segnet with a decent 30-time DCDO for proper statistics costs a nominal 280 milliseconds, our one-shot approach takes 12.8 milliseconds, much more efficient and amicable to online applications.
Conclusions: our approach is able to achieve comparable error prediction power as the popular segnet implementation with a much smaller computational cost. This advantage is desirable in real-time, online applications, or interactive procedures. It is also compatible to any error-driven segmentation refinement procedure using segnet.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
