Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation
Authors: Szymon Płotka, Gizem Mert, Maciej Chrabaszcz, Ewa Szczurek, Arkadiusz Sitek
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through comprehensive experiments on four publicly available datasets, including PANORAMA [2], AMOS [23], Fe TA 2022 [37], and MVSeg [10] as well as one in-house CT dataset, we demonstrate that Mamba-Ho ME outperforms current state-of-the-art methods in both segmentation accuracy and computational efficiency, while generalizing effectively across three major 3D medical imaging modalities: CT, MRI, and US. |
| Researcher Affiliation | Academia | 1 Faculty of Mathematics, Informatics, and Mechanics, University of Warsaw, Poland 2 Faculty of Mathematics and Computer Science, Jagiellonian University, Poland 3 Institute of AI for Health, Helmholtz Munich, Germany 4 Faculty of Electronics and Information Technology, Warsaw University of Technology, Poland 5 NASK National Research Institute, Poland 6 Faculty of Radiology, Massachusetts General Hospital, USA 7 Department of Radiology, Harvard Medical School, USA |
| Pseudocode | No | The paper describes the methodology and architecture using text and diagrams (Figure 1), but it does not contain any explicitly labeled "Pseudocode" or "Algorithm" blocks. |
| Open Source Code | Yes | The code is publicly available at github.com/gmum/Mamba Ho ME. |
| Open Datasets | Yes | Through comprehensive experiments on four publicly available datasets, including PANORAMA [2], AMOS [23], Fe TA 2022 [37], and MVSeg [10] as well as one in-house CT dataset, we demonstrate that Mamba-Ho ME outperforms current state-of-the-art methods in both segmentation accuracy and computational efficiency, while generalizing effectively across three major 3D medical imaging modalities: CT, MRI, and US. |
| Dataset Splits | Yes | Pre-training. ...data are partitioned into training and validation sets using a stratified 85:15 split to maintain a balanced distribution of CT and MRI modalities. (...) PANORAMA. ...We employ a stratified 80:20 split, resulting in 1,571 training and 393 validation scans (...) AMOS. ...The original training set is partitioned into training and validation subsets using a randomized 80:20 split. (...) Fe TA 2022. ...We evaluate model performance using a 5-fold cross-validation strategy (...) MVSeg. ...We partition the dataset into independent training, validation, and testing sets consisting of 105, 30, and 40 scans, respectively. |
| Hardware Specification | Yes | The experiments were conducted on a workstation equipped with 8 NVIDIA H100 GPUs. (...) For hardware, we employed an NVIDIA DGX system equipped with 8 NVIDIA H100 80 GB GPUs. We used all eight GPUs for pre-training, while training, fine-tuning, and evaluation were performed on a single NVIDIA H100 80 GB GPU. |
| Software Dependencies | Yes | For implementation, we employ Python 3.11, Py Torch 2.4 [35], and MONAI 1.3.0 within a Distributed Data-Parallel (DDP) training setup. |
| Experiment Setup | Yes | All training is performed using the LDice CE loss function and the Adam W optimizer, with an initial learning rate of 1e-4 controlled by a cosine annealing scheduler [32], a weight decay of 1e-5, and a batch size of 2. All models are trained with 32-bit floating-point precision to ensure numerical stability and to standardize the training process across all experiments. A detailed description of the implementation can be found in Appendix B. |