Work Description

Date: 25 February, 2024

Dataset Title: Deep Learning (DL) segmentation tools for murine tibia bone marrow from 3D MRI

Dataset Creators: Dariya Malyarenko, Humera Tariq, Aman Kushwaha, Rami F Mourad, Kevin Heist, Thomas L Chenevert, Brian D Ross, Heang-Ping Chen, Lubomir Hadjiiski

Dataset Contact: Dariya Malyarenko, dariya@umich.edu

Funding: NIH

Overview:
...
The 3D GRE MRI data for murine model of myelofbifrosis with expert segmentations of mouse tibia was used to train Attention UNET model to automate bone marrow segmentation for measurements of imaging biomarkers. This dataset consists of three archives:
(1) containing the source MRI images in Meta-image-header (MHD) format with resulting segmentation labels by two experts and four UNET models with different training scenarios;
(2) corresponding training models; and
(3) deep-learning (DL)-based segmentation tools for application to future murine tibia MRI data.
The MHD images are an ITK compatible format that can be viewed in standard image viewer, like 3D Slicer. The image archive is structured with a directory tree that contains \"mouseID"\"scan-date"\"segmentaion-scenario"\. The "training model" archive containes DL-model labeled by the data subset, and "deployment" archive containes the DL-segmentation software.

Methodology:
...
A murine model of myelofibrosis in tibia was used in a co-clinical trial to evaluate segmentation methods for application of image-based biomarkers to assess disease status. All animal procedures were approved by the University of Michigan Institutional Animal Care and Use Committee (IACUC PRO:00006795, approved 01/21/2016) and were performed in compliance with all relevant ethical regulations therein. The 3D MRI dataset (32 mice with 157 scans including 49 test–retest pairs scanned on consecutive days) was split into approximately 70% training, 10% validation, and 20% test subsets. Two expert annotators (EA1 and EA2) performed manual segmentations of the mouse tibia (EA1: all data; EA2: test and validation). Attention U-net (A-U-net) model performance was assessed for accuracy with respect to EA1 reference for four training scenarios: full training (TS), two half-splits (TS1,TS2), and a single-mouse (TSM) subsets. The repeatability of computer versus expert segmentations of test–retest pairs was measured by within-subject coefficient of variance (%wCV). Details of the segmentation model training and perfromance are provided in PMC10037585 at https://doi.org/10.3390/tomography9020048.

Date Coverage: 2020-08-24 to 2022-12-15

Instrument and/or Software specifications: See related publication for details. The software is deployed as a pre-compiled 64-bit Window-10-exe that uses Pythone DL libraries.

Files contained here:
-"README_tibiaSegR1.txt" - instructions for deployment of DL segmentation software (see below).

-"PreclinicalMyelofibrosisModel_MouseTibia.pdf" - animal protocol description

-"MFtibiaSegData.zip" contains 3D MRI dataset for 32 mice with 157 imaging sessions including 49 test–retest pairs scanned on consecutive days organized by mouse ID, image date and segmentation model (also availbale from https://github.com/dumichgh/MFJK1_Segmentation_MHDs). "SegData" archive has a directory tree with MHD images and segmentations by expert arranged as:
\"mouseID"\"scan-date"\"seg_scenario"\. The segmentaion scenarios are labeled as :
"EA1" -- manual expert annotator 1;
"EA2" -- manual expert annotator 2;
"TS" -- attention UNet for full training set;*
"TS1" -- attention UNet for "Split1" training set;
"TS2" -- attention UNet for "Split2" training set;
"TSM" -- attention UNet for single-mouse training;

*(Note that EA2 annotations were NOT performed for animals in the Training set.)

-"tibiaSeg_deploy.zip" contains the DL segmentation software. Once unzipped you will find the directory structure with pre-compiled Windows executables for Pytorch-based Attention-UNET segmentation and set-up files. The deployment steps are described in the included "READme".

-"train_model_bak.zip" includes the segmentation models built for training scenarios. When unzipped you will find the archived DL-models for "TS" (model 66); "TS1" (model 65); "TS2" (model 47) and "TSM" (model 48).(The full-model 'model66.pth' is trained on 107 scans, 23 mice, 32 test-retest pairs annotated by a single expert).

-"MFJK1_seg_data_info.xlsx" contains assignment of the scans to the training, validation and test sets, while

-"MFJK1_tibia_seg_vol_RepeatabilityAnalysis.xlsx" summarizes test-retest results for different segmentation and training scenarios.

-"MFJK1_tibia_seg_tabular_files.zip" contains .csv-converted versions of "MFJK1_seg_data_info.xlsx" and "MFJK1_seg_data_info.xlsx" including each sheet as a .csv file, formulas in .txt format, and a visual representation of each sheet in .html format.

Related publication(s):
Kushwaha A, Mourad RF, Heist K, Tariq H, Chan HP, Ross BD, Chenevert TL, Malyarenko D, Hadjiiski LM. Improved Repeatability of Mouse Tibia Volume Segmentation in Murine Myelofibrosis Model Using Deep Learning. Tomography. 2023 Mar 7;9(2):589-602. doi: 10.3390/tomography9020048. PMID: 36961007; PMCID: PMC10037585.

Use and Access:
This data set and software are made available under a Creative Commons Attribution-Noncommercial (CC-BY-NC 4.0) license.

To Cite Data:
Dariya Malyarenko, Humera Tariq, Aman Kushwaha, Rami F Mourad, Kevin Heist, Thomas L Chenevert, Brian D Ross, Heang-Ping Chen, Lubomir Hadjiiski Deep Learning (DL) segmentation of murine tibia bone marrow from 3D MRI [Data set], University of Michigan - Deep Blue Data.