Deep Learning in Medical Image Registration: Magic or Mirage?

Authors: Rohit Jena, Deeksha Sethi, Pratik Chaudhari, James Gee

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This hypothesis is thoroughly validated with state-of-the-art classical and learning-based methods.
Researcher Affiliation Academia Rohit Jena1,4 Deeksha Sethi1 Pratik Chaudhari1,2, James C. Gee1,3,4, 1Computer and Information Science 2Electrical and Systems Engineering 3Radiology 4 Penn Image Computing and Science Laboratory {rjena, deesethi, pratikac}@seas.upenn.edu, gee@upenn.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes We make all evaluation scripts and trained models public2 to encourage fairness and transparency in evaluations. 2https://github.com/rohitrango/Magic-or-Mirage/
Open Datasets Yes We consider four brain datasets OASIS, LPBA40, MGH10, and IBSR18
Dataset Splits Yes We split the OASIS dataset into a training set of 364 images and a validation set of 50 images.
Hardware Specification Yes All experiments are run on a cluster with 2 AMD EPYC 7713 CPUs and 8 NVIDIA A6000 GPUs.
Software Dependencies No The paper lists various state-of-the-art classical and DLIR baselines (e.g., ANTs, Nifty Reg, Synth Morph, Trans Morph) but does not specify software versions for any of them or for any other key dependencies.
Experiment Setup No For all DLIR methods, we use pretrained models if they are trained with Eq. (1), or train them with the architecture and hyperparameters provided in their original source code. [...] For all classical methods, we follow their recommended hyperparameters and run till convergence.