270 likes | 405 Views
Neuronal Reconstruction Workshop. Darren R. Myatt*, Slawomir J. Nasuto, Giorgio A. Ascoli. *d.r.myatt@reading.ac.uk , http://www.rdg.ac.uk/neuromantic. More Acknowledgements. Thanks also go to Tye Hadlington Nathan Skene Kerry Brown (GMU)
E N D
Neuronal Reconstruction Workshop Darren R. Myatt*, Slawomir J. Nasuto, Giorgio A. Ascoli. *d.r.myatt@reading.ac.uk, http://www.rdg.ac.uk/neuromantic
More Acknowledgements • Thanks also go to • Tye Hadlington • Nathan Skene • Kerry Brown (GMU) • Thanks specifically do not go to the Heathrow Airport Security team
Requirements for this workshop • Laptop running/emulating Windows • WINE should be ok, except for possibly the 3D display • A reasonable amount of RAM • 1 Gig recommended, although 512M will be OK – less is possible, but not great • A standard 3 button mouse/trackball with mouse wheel • Not strictly necessary but strongly preferable – I have a few spares to hand out • Either a working CD-ROM drive or USB port that will recognise a flash drive • If you have neither of these, then I will begin to suspect that you are in league with the Heathrow airport security team in making my life more difficult than it needs to be
Workshop Aims • Provide participants with direct experience of reconstructing neurons and the challenges involved in resolving ambiguities • Give a tutorial with the freeware Neuromantic application • Semi-manual reconstruction • Semi-automatic reconstruction • To generate discussion about best practice for reconstructing dendritic trees • Consistency remains a problem • Gather feedback and recommendations on improvement for the Neuromantic tool • The workshop length is not set in stone but will probably last for around two hours
Why Reconstruct Neurons? • Allows the validation and refinement of simulations of neuronal behaviour • Compare between simulation (via NEURON or GENESIS) and electrophysiological testing • Gaining large enough populations of reconstructed neurons allows insight into the morphological variation observed in each class. • Facilitates the identification of dendritic abnormalities associated with brain disease • Epilepsy, Alzheimer’s disease, some forms of retardation etc. • Compare statistical properties of trees between control and experimental conditions (via L-Measure, for example)
Is it Live or is it Memorex? • Two main options for reconstruction… • Live imaging (NeuroLucida) • Advantages: no real memory requirement, no discretisation in Z. • Disadvantages: specimen degradation over time and Z drift on stage • Reconstruction from an image stack • Advantages: minimal specimen degradation and Z drift • Disadvantages: can require large amounts of storage and Z values are usually discretised. • A motorised stage is strongly preferred.
Flavours of Reconstruction • Reconstruction methods may be split into 4 (or possibly 5) broad classes • Manual • Semi-manual • Semi-automatic • Automatic • So automatic that you don’t even need to turn up to work any more
Manual Reconstruction • User has to do define every neurite compartment with very little or no assistance • Incredibly laborious and time consuming • Camera Lucida • Pencil and paper tracing via a system of prisms (it still exists!) • Neuron_Morpho • Freeware plug-in for ImageJ • Original inspiration for Neuromantic
Semi-manual Reconstruction • Each segment is still added manually by the user • Application gives some assistance in some elements of the task to reduce effort e.g. auto focussing, useful visualisation • NeuroLucida (without AutoNeuron), Neuromantic on manual mode • Generally considered to be the most accurate method of reconstruction, but still highly time consuming
Semi-automatic Reconstruction • Application requires constant user-interaction, but the application requires mainly topological information. • Define beginning and end points of a dendrite, and the neurite is traced out automatically • NeuronJ • Freeware plug-in for ImageJ (single image only) • Derived from the robust LiveWire algorithm • Neuromantic • Semi-auto tracing is a 3D extension of the NeuronJ algorithm with post-processing • Also includes radius estimation
Automatic Reconstruction • What everybody really wants… • Current automatic techniques are generally limited to high quality microscopy data (e.g. confocal fluorescence) • AutoNeuron for NeuroLucida, NeuronStudio • Numerous skeletonisation techniques, and also the Rayburst algorithm. • The outputs frequently require cleaning up to bring reconstruction accuracy up to the required standard
Which flavour to choose? • t(Automatic)+t(Clean Up)<t(Manual)? • Realistically, the clean up time will always be non-zero, except in trivial cases • With noisy data, fully automatic reconstruction is unlikely to be possible • A good reconstruction application should • make it as easy as possible to spot errors • have good manual editing capabilities to facilitate clean up
Issues with reconstruction • Interuser/Intrauser variation… • Different users on the same system • The same user on different systems • Even the same user reconstructing the same neuron on the same system! • Thin dendrites (relative to image resolution) are a particular problem, as errors in radius estimation can have a large impact on surface area and cross-sectional area. • Increased automation should increase consistency, but accuracy may still be a problem.
Example from Jaeger, 2001 • These reconstructions were performed in NeuroLucida by experienced users • Surface area range shows over 20% variation, which has a lot of implications for behavioural simulations • and this is just variation over individual dendrites, not a whole dendritic tree!
Pyramidal Neuron Example • All 10 participants were complete novices at neuronal reconstruction • Interquartile range of surface area shows around 15% variation • Interquartile range of volume is around 30% variation • Includes thicker neurites as well as thin
Neuromantic • Freeware application for making 3D reconstructions of neurons from serial image stacks • Programmed in C++ Builder • Can function on any form of microscopy data from non-deconvolved widefield stacks upwards. • Semi-manual tracing • Manually position new compartments, which may then be edited afterwards as necessary • Semi-automatic tracing • Longer neurite sections can be traced out automatically, and the radius is calculated at each point • The neuron can also be visualised in 3D to help identify and correct errors
Basic Interface Mode Buttons Mode options Overlaid Reconstruction Image Stack Stack Bar Image Processing
Installation Time! • CD/Flash drive contains • Neuromantic directory • Stack containing basal tree of a pyramidal neuron • Simply copy the Neuromantic directory onto your computer somewhere, and it should be fine (hopefully!) • Copy the stack to a directory nearby • Run the Neuromantic executable V1.4.1 to make sure everything is working
Getting Started • An updated manual may be found in Manual.pdf in the Neuromantic directory • Load in the stack by pressing F2 or File->Load Stack and selecting the first image • Wait for a while under the stack loads (it’s 387 Megabytes in total with 86 images) – the status bar shows the current progress • Halve stack size if you are forced to use virtual RAM otherwise (Options->Stack->Halve Stack Size)
Stack Navigation • Most functionality is always present on the mouse for speed • Drag the stack around with the right button • Zoom in/out by rolling the mouse wheel (or -/+ keys for those without) • Use the stack bar or hold down the middle mouse button and move vertically to scroll through the different images (z axis) • Middle clicking the mouse button auto-focuses at that position (+/- 5 slices) • Hold SHIFT while middle clicking to auto-focus over all images
Semi-manual Reconstruction • Each compartment is added by dragging a line from one edge of the dendrite to the other, thus providing an estimate of the radius • The compartment added is of the type defined by the radio buttons in the Manual panel to the right • Every time a new compartment is added its parent is set to the currently selected compartment • So add a compartment, then auto-focus on the next position down the dendrite, then add the next etc. • In order to create a branch point, select the desired compartment with a left mouse click, then carry on as before
Selecting Compartments • As you move the cursor towards the centre of a compartment it will change, indicating that you can manipulate that segment • Left click a compartment to select it • SHIFT whilst selecting to add to the current selection • CTRL whilst selecting to select an entire branch • ALT to select all the compartments of the same type • CTRL+I inverts the current selection • CTRL+D deselects all compartments • Using these controls it is possible to efficiently select any set of compartments, such as a subtree.
Editing Compartments • Selected compartments can be dragged around in the x/y plane using the left mouse button • The Z value is altered by selecting a compartment, navigating to the new desired image slice, and then pressing CTRL+C (or Edit->Set Z To Current Slice) • The radius of a compartment is altered by holding down CTRL, and dragging with the middle button • Press DELETE to delete all selected compartment
Semi-automatic Reconstruction • Newly added to the application • Still a bit of a Work In Progress, as it is not as intuitive as I would like yet • Employs an extension to 3D of the semi-automatic algorithm used in NeuronJ • Includes estimate of dendritic radius • Additional post-processing to improve accuracy
Semi-automatic Reconstruction • Employs Steerable Gaussian Filters to perform the image processing • Efficiently yields information on the position of neurites and flow direction from eigen analysis of the Hessian matrix • The standard deviation of the Gaussian determines the radius of the neurites detected • A graph search (via Djikstra’s algorithm) is then performed to calculate the optimal route via the defined cost function
Patchwork Method • Pre-processing on the entire image stack is expensive in both time and space. • For the basal stack used in this workshop, around 10Gigabytes of RAM would be required • Therefore, to avoid this issue, only the necessary patches of the image are image processed and routed.
Conclusions • Discussed reconstruction in general and some of the challenges associated with it • Given participants experience of the Neuromantic application, in terms of both its semi-manual and semi-automatic capabilities • I hope you have enjoyed yourselves!