In neuroconductor-devel/mimosa: 'MIMoSA': A Method for Inter-Modal Segmentation Analysis

mimosa

mimosa is an R package based on the paper: MIMoSA: An Automated Method for Inter-Modal Segmentation Analysis of Multiple Sclerosis Brain Lesions by Valcarcel et al. published in the Journal of Neuroimaging. This package creates data structures necessary for training and testing and then allows the user to train a model and then apply the trained model to generate probability maps and predicted lesion segmentations.

Installation

To install the package from neuroconductor, type:

source("https://neuroconductor.org/neurocLite.R")
neuro_install("mimosa")

To get the latest development version from GitHub:

devtools::install_github('avalcarcel9/mimosa')

avalcarcel9 badges:

muschellij2 badges:

Vignette

For a full implementation of the methods with output please see our vignette.

Functions

Below is a list of the functions and a description of options available to utilize through the mimosa package.

count_stats

This function calculates true positive rate, false positive rate, false negative rate, false positive count, and sensitivity.

Formulas for how these are calculated are provided in the 'Evaluate Performance' section.

count_stats(gold_standard, 
            predicted_segmentation, 
            k, 
            percent_overlap = NULL, 
            verbose = TRUE)

Arguments

• gold_standard Gold standard segmentation mask of class nifti
• predicted_segmentation Predicted segmentation mask volume of class nifti
• k Minimum number of voxels for a segmentation cluster/component
• percent_overlap Proportion of gold standard segmentation to be overlapped by predicted
• verbose Logical indicating printing diagnostic output

mimosa_data

This function creates the training vectors from a single MRI study that has FLAIR, T1, T2, and PD volumes. When utilizing the function for training you will also need gold standard binary segmentation mask. The function can create a tissue mask (or the user can supply a brain mask), a binary mask of candidate voxels for lesion segmentation, smoothed volumes, and coupling maps. The user may supply already normalized data if they wish to use an alternative normalization method.

mimosa_data(brain_mask, 
            FLAIR, 
            T1, 
            T2 = NULL, 
            PD = NULL, 
            tissue = FALSE, 
            gold_standard = NULL, 
            normalize = 'no', 
            cand_mask = NULL, 
            slices = NULL, 
            orientation = c("axial", "coronal", "sagittal"), 
            cores = 1, 
            verbose = TRUE)

Arguments

• brain_mask brain or tissue mask of class nifti
• FLAIR volume of class nifti
• T1 volume of class nifti
• T2 volume of class nifti. If not available use NULL.
• PD volume of class nifti. If not available use NULL.
• tissue is a logical value that determines whether the brain mask is a full brain mask or tissue mask (excludes CSF), should be FALSE unless you provide the tissue mask as the brain_mask object
• gold_standard gold standard lesion segmentation mask of class nifti
• normalize by default is 'no' will not normalize images. To normalize images use inputs 'Z' for z-score normalization slices vector of desired slices to train on, for WhiteStripe use 'WS'
• cand_mask is NULL to use candidate mask procedure proposed with method or a nifti object to be used as the candidate mask
• slices vector of desired slices to train on, if NULL then train over the entire brain mask
• orientation string value telling which orientation the training slices are specified in, can take the values of "axial","coronal", and "sagittal",
• cores 1 numeric indicating the number of cores to be used (no more than 4 is useful for this software implementation)
• verbose logical indicating printing diagnostic output

mimosa_training

This function trains the MIMoSA model from the data frames produced by mimosa_data on all subjects and determines an optimal threshold based on training data.

mimosa_training(brain_mask, 
                FLAIR, 
                T1, 
                T2 = NULL, 
                PD = NULL, 
                tissue = FALSE, 
                gold_standard, 
                normalize = 'no', 
                slices = NULL, 
                orientation = c("axial", "coronal", "sagittal"), 
                cores = 1, 
                verbose = TRUE, 
                outdir = NULL, 
                optimal_threshold = NULL)

Arguments

• brain_mask vector of full path to brain mask
• FLAIR vector of full path to FLAIR
• T1 vector of full path to T1
• T2 vector of full path to T2 if available. If not use NULL.
• PD vector of full path to PD if available. If not use NULL.
• tissue is a logical value that determines whether the brain mask is a full brain mask or tissue mask (excludes CSF), should be FALSE unless you provide the tissue mask as the brain_mask object
• gold_standard vector of full path to gold standard segmentations. Typically manually segmented images.
• normalize by default is 'no' will not normalize images. To normalize images use inputs 'Z' for z-score normalization slices vector of desired slices to train on, for WhiteStripe use 'WS'
• orientation if NULL then train over the entire brain mask orientation string value telling which orientation the training slices are specified in, can take the values of "axial", "sagittal", or "coronal"
• cores numeric indicating the number of cores to be used (no more than 4 is useful for this software implementation)
• verbose logical indicating printing diagnostic output
• outdir vector of paths/IDs to be pasted to objects that will be saved. NULL if objects are not to be saved
• optimal_threshold = NULL. To run algorithm provide vector of thresholds

Pre-Trained Models

The method performs best when trained on data. Since gold standard manual segmentations are not always delineated though through the mimosa package we have trained four (4) distinct models for use. These models can be called and are stored under the following names:

• mimosa_model trained using FLAIR, T1, T2, and PD imaging modalities
• mimosa_model_No_PD trained using FLAIR, T1, and T2 imaging modalities
• mimosa_model_No_T2 trained using FLAIR, T1, and PD imaging modalities
• mimosa_model_No_PD_T2 trained using FLAIR and T1 imaging modalities

Users should use the model which matches their imaging sequence. For example, if you only have data collected for FLAIR and T1 modalities then you should use the mimosa_model_No_PD_No_T2.

mimosa_model
mimosa_model_No_PD
mimosa_model_No_T2
mimosa_model_No_PD_No_T2

neuroconductor-devel/mimosa documentation built on Sept. 21, 2019, 6:24 p.m.