Benchmarking Encoder-Decoder Architectures for Biplanar X-ray to 3D Bone Shape Reconstruction

Mahesh Shakya, Bishesh Khanal

Keywords: 2D-3D Reconstruction, Object Reconstruction, Medical Applications, Encoder-Decoder Architectures

Abstract:

Various deep learning pipelines have been proposed for 3D Bone Shape Reconstruction from Biplanar X-rays. Although these methods individually report excellent results, we do not know how these architecture pipelines compare against each other since they are reported on different anatomy, datasets, and cohort distribution. We benchmark these disparate architectures on equal footing on three different anatomies using public datasets. We describe various benchmarking tasks to simulate real-world clinical scenarios including reconstruction of fractured bones, bones with implants, robustness to population shift, and estimate clinical parameters.

We provide an open-source implementation of SOTA architectures, dataset pipelines, and extraction of clinical parameters.
Comparing the encoder-decoder architectures with baseline retrieval models, we find that the encoder-decoder methods are able to learn from data and are much better than retrieval baselines. However, the best methods have limited difference on performance, but the domain shift plays an important role in deteriorating the performance of these methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )