Cranial Implant Prediction by Learning an Ensemble of Slice-Based Skull Completion Networks

Abstract

The development of automatic skull reconstruction methods has dramatically reduced the time and expense to repair skull defects. In this study, an ensemble-learning-based method is proposed for skull implant prediction. To overcome the potential overfit problem in 3-D volume analysis using deep learning, a set of 2-D defective skull images is generated by slicing 3-D volumes along the X, Y, and Z axes. We further introduce an RNN model in this method to compensate for the loss of global skull information in the 2-D implant prediction. Over the implant estimation problem in Task 1 of the AutoImplant 2021 challenge, we observe a considerable performance boost from our averaging ensemble strategy and noise removal filtering. The codes for our method as well as our pretrained models is accessible with https://github.com/YouJianFengXue/Cranial-implant-prediction-by-learning-an-ensemble-of-slice-based-skull-completion-networks.

Appendix

In this challenge, we made several submissions for algorithm assessment and improvement. The specific quantitative evaluation of our submissions that comes from the challenge organizer are presented in this section. Specifically, Table 2 reports the performance of the baseline CNN model on the test samples in the five folders. Note that our submission encountered unexpected errors over the test samples in the folder of random2. So we particularly include this information here for your reference or any further research and report the performance of the baseline through two sets of numerical values in Table 1, where the values in the first row are computed from the results of the first four folders and the second line corresponds to the performance assessment over all 5 folders. We believe that quantitative measurement in the first row of Table 1 reflects the the performance of our baseline model. Similarly, Table 3 and Table 4 report the specific numerical results for our later submissions. Slightly different from the baseline model, the final results presented in Table 1 are averaged over the five folders.

Table 2. The quantitative results of the baseline network from the challenge organizer. Note that our submission had unexpected errors on the test samples in the folder of random2. So we faithfully mark the error here with the star sign $*$ for your reference.

Table 3. The quantitative results of the our CNN+RNN models from the challenge organizer.

Table 4. The quantitative evaluation of the entire solution from the challenge organizer.