LeopardGecko

This package contains tools for 3D semantic segmentation.

segmentor.py

LeopardGecko segmentation is powered by volume-segmantics, which makes training and predictions along multiple-axis, resulting in multiple predictions. LeopardGecko collects these predictions and fuse them together with a MLP classifier.

segmentor.py has training and routine functions, please see examples in the scripts folder.#

In addition, it can caclulate the Consistency Score of the predicted volume, which is a sort of a volumetric heatmap of how consistent the multiple predictions were for each voxel. and can be used as a metric that measures the certainty of the prediction.

metrics.py

Some metrics functions from the predicted data

AvgPool3D_LargeData()

Uses pytrorch to do a 3D average pooling, by providing parameters data3d (data volume), w_width ( width in pixels of the cube chunk that is calculated in the GPU), k_width (does the average in the cube with this width), s_stride (step, in pixels for next average calculation).

This operation is normally done with the weighted predicted data to determine regions with higher consistency.

Score data

Class to store and analyse score data. Contains routines to generate histograms and suggest volumes of interest based in histogram positions.

Class PredictedData

Constructor (filename): initialise with a hdf5 containing the volume data that has been predicted and combined from several different orientations. This file is normally the output from the prediction produced by binary unet segmentation https://github.com/rosalindfranklininstitute/UnetSegmentation

lg-genpredcsreport.py

Utility that can be used to generate a report of consistency score analysis from a combined prediction file. This script uses lg-genpredreport.ipynb notebook. Information about how to use and parameters to use can be obtained by running python lg-genpredcsreport.ipynb --help

(old) AvgPooling3DConsistencyData.ipynb

An average pooling operation using PyTorch CUDA routines are used. The data obtained gives a score about how consistent the data is in the each of smaller volumes.

(old) AnalyseAvgPoolResults.ipynb

Data from the average pooling is analysed and consistency score based visualisations are available