Download presentation
Presentation is loading. Please wait.
Published byElfreda Potter Modified over 6 years ago
1
Watershed Based Visualization of High-Density EEG Coherence
Michael ten Caat1,2, Natasha M. Maurits2,3, Jos B.T.M. Roerdink1,2 1. Institute for Mathematics and Computing Science, University of Groningen 2. BCN Neuroimaging Center, University of Groningen 3. Department of Neurology, University Medical Center Groningen ISMM, Rio de Janeiro, October 2007
2
Outline Introduction electroencephalography (EEG) coherence
Data-driven visualization: Existing: maximal clique based (MCB) Novel: watershed based (WB) Comparison MCB and WB Conclusion, Discussion, Questions
3
EEG coherence EEG measures electrical brain activity using up to 512 electrodes Synchronous electrical activity between brain regions is assumed to imply functional relationships between these regions Measure for this synchrony: EEG coherence Coherence: similarity of a pair of signals, as a function of frequency (values between 0 and 1) Generally computed per frequency band
4
Data Participant listened to high (15%) and low (85%) tones
High tones were counted 20 brain responses to high tones were recorded in 20 segments, using 128 EEG electrodes (high-density EEG) Significance threshold for coherence: θ = 1 - p1/(L-1) (p = probability, L = no. segments) Here, for L = 20 and p = 0.05 : θ = 0.15 High-density EEG refers to EEG recorded with large number of electrodes.
5
Conventional data-driven visualization
1-3 Hz significance threshold coh > 0.15 top 10% coh > 0.37 top 1% coh > 0.91 All electrodes and all significant coherences Result: visually cluttered edges
6
Data-driven ROI: Functional Unit (FU)
1-3 Hz Functional Unit (FU): spatially connected clique (clique = maximally connected subgraph) ROI Volume is the motivation to define a functional unit as a spatially connected clique. Volume conduction: activity from one source recorded by multiple electrodes [Holsheimer & Feenstra, 1977]
7
Functional Unit Map
8
Functional Unit Map
9
1. Define neighborship electrodes
Voronoi diagram of (planar projection of) electrode positions For example, the blue vertex surrounded by fourVoronoi-edges (red) has four spatial Voronoi neighbors (red) For each point in a cell, the closest electrode is the electrode in that cell
10
2. Find spatially connected cliques
1-3 Hz Larger ROIs are assumed to correspond to stronger source signals and are therefore considered to be more interesting. => Detect spatially connected maximal cliques. Functional unit: spatially connected maximal clique
11
3. Determine functional units
FU visualized as connected set of cells with identical gray value 1-3Hz
12
3. Determine functional units
FU visualized as connected set of cells with identical gray value 1-3Hz The sum of all coherences in a Voronoi-connected maximal clique is referred to as total strength. a - The Voronoi-connected maximal cliques are sorted by their total strength, from high to low. b - The vertex set corresponding to the one with the highest total strength is labeled W1. c - The vertices in W1 are assigned the label 1 and are removed from the remaining (maximal) cliques. d - If the remaining cliques are not Voronoi-connected any more, they are split into Voronoi-connected components. e - For every changed clique the total strength is recomputed, and the clique is inserted in an appropriate location in the sorted list of cliques. This procedure (a - e) is repeated, until all vertices received a label, or until all maximal cliques have been considered. This results in vertex sets labeled W_1, ...,W_M. Every vertex set W_i (i 2 {1, 2, ...,M}) is considered to be a functional unit, being a clique consisting of Voronoi-connected vertices.
13
3. Determine functional units
FU visualized as connected set of cells with identical gray value Adjacent FUs different gray values Geographic center of an FU indicated by cross inside circle 1-3Hz
14
4. Show only FUs larger than 5 cells
1-3Hz
15
5. Inter-FU coherence Calculate average coherence S between FUs
Lines connect FUs if S > significance threshold Color mapped to value S 1-3Hz
16
Time complexity Because maximal clique detection has time complexity O(3^{n/3})
17
Goal Because maximal clique detection has time complexity O(3^{n/3})
18
Watershed Based FU Maps
Initialization: Define coherence of vertex v as average coherence between v and its Voronoi neighbors Select all local maxima as markers Label markers and associate markers with basins Initialize edge queues with edges between markers and their spatial neighbors (i.e. Voronoi neighbors) Sort edges in decreasing order of value Extend basins as follows, see: Illustration Pseudocode
19
Watershed Based FU maps
basin B and Voronoi-neighbors queue: edges between B and Voronoi neighbors candidates connected to all w in B Electrode (Imaginary example) Left: blue rectangles are basin vertices, green crosses are Voronoi neighbors of this basin (4-connected) Middle: goldish lines are edges between basin vertices and Voronoi neighbors (put in a sorted edge queue from high to low values) Right: red rectangles are candidates that may extend the basin; imagine that every red rectangle is connected each of the blue rectangles in the underlying graph. B = basin (clique) V = {v | v is Voronoi neighbor of a basin vertex} C = candidates = { v | v connected to all w in B} queue = {ei(vi,wi) = | vi = , wi = : ek > em iff k < m}
20
Watershed Based FU maps
dequeue first edge from ‘queue’ ; w not in C dequeue first edge; w in C add w to B w w w Electrode Left: return and remove the first edge from the queue (i.e. the edge with the highest value) between a blue rectangle and a green cross. The vertex corresponding to the green cross is not in the candidates set. Nothing happens. Middle: the next edge is between a basin vertex (blue rectangle) and a candidate vertex (red rectangle). The candidate vertex (w) is B = basin (clique) V = {v | v is Voronoi neighbor of a basin vertex} C = candidates = { v | v connected to all w in B} queue = {ei(vi,wi) = | vi = , wi = : ek > em iff k < m}
21
Watershed Based FU maps
add Voronoi neighbors v of w to V ; add e(w,v) to queue remove vertices not connected to w from C w w w Electrode Left is same as last ‘graph’ from previous slide. Middle: edges between w and spatial neighbors of w (green crosses) are added to queue (goldish lines) Right: imagine that some red rectangles are not connected to w in the underlying graph, and remove these. B = basin (clique) V = {v | v is Voronoi neighbor of a basin vertex} C = candidates = { v | v connected to all w in B} queue = {ei(vi,wi) = | vi = , wi = : ek > em iff k < m}
22
Watershed Based FU Maps
1 while queue not empty do 2 e(v,w) ← dequeue(queue) (* v in basinL(v) *) 3 if L(w) = 0 (* w not in any basin *) 4 if w in candidatesL(v) then 5 candidatesL(v) ← intersect(candidatesL(v),adjcohw) 6 L(w) ← L(v) 7 basinL(v) ← insertVertexSort(w,basinL(v)) 8 for all v* in adjvorw do if L(v*)=0 then insertEdgeSort(e(w,v*),queue) end if 10 end for 11 end if 12 end if 13 end while basin_i vertices in basin (sorted by vertex nr) e(v,w) edge between vertex v and w L(v) basin label of vertex v adjcohw vertices adjacent to w in the coherence graph (sorted by nr) adjvorw Voronoi neighbors of w (sorted by nr) queue list of edges sorted in decreasing order; e(v,w) is added v is put into a basin for every unlabeled Voronoi neighbor w of v As long as the edge queue is not empty Return and remove the first edge e(v,w) from the queue Vertex v was labeled when this edge was inserted, w was still unlabeled. If w still unlabeled (i.e. w still not in any basin) then If w is connected to each basin vertex then Intersect the candidates with the neighbors of w (in the coherence graph) Assign to w the same basin label as v Add w to the basin which contains v For each Voronoi neighbor v* of w If v* still unlabeled (i.e. not in any basin yet) then add e(w,v*) to the queue.
23
Watershed Based FU Maps
initialization queue: O(n2 log n) 1 while queue not empty do 2 e(v,w) ← dequeue(queue) (* v in basinL(v) *) 3 if L(w) = 0 (* w not in any basin *) 4 if w in candidatesL(v) then 5 candidatesL(v) ← intersect(candidatesL(v),adjcohw) 6 L(w) ← L(v) 7 basinL(v) ← insertVertexSort(w,basinL(v)) 8 for all v* in adjvorw do if L(v*)=0 then insertEdgeSort(e(w,v*),queue) end if 10 end for 11 end if 12 end if 13 end while O(m) = O(n2) : m = # edges; n = # vertices = # electrodes O(log n) : binary search in sorted set candidates O(n) : intersection two vertex sets of size at most n O(n log m) = O(n log n) : insert at most n edges in sorted edge list at most n times } Initialization of queue: occurs before the pseudocode shown here. Its time complexity is O(n^2\log{n}). Main loop (is this pseudo code): Total time complexity at first sight O(n^3\log{n}). Last three steps executed at most n times. Total time complexity actually O(n^2\log{n}). Total time-complexity: O(n2 log n)
24
FU Map: maximal clique based, watershed based method
1-3 Hz Dataset max. clique based water- shed based Comparison. First 1-3Hz frequency band. Similarity between FU map for MCB and WB method (per dataset) Similar FU locations Similar inter-FU connections Largest difference for dataset 2: larger number of inter-FU connections for WB method – come back to that in next slide.
25
FU Map: maximal clique based, watershed based method
13-20 Hz FU Map: maximal clique based, watershed based method 1-3 Hz, dataset 2 max. clique based, p=.05 Comparing the middle with the left FU map, the middle one shows over-segmentation (accepting the MCB as the gold standard). Six inter-FU connections for MCB method for p=.05 (left) similar to inter-FU connections with higher value for WB method (see right). Even though adapting a threshold is not the fanciest you can do, this indicates that the inter-FU connections with the highest value (i.e. the most important ones) correspond between the MCB and the WB method. (The discussion will mention a different solution.) watershed based, p=.05 watershed based, p=.001 ‘oversegmentation’
26
FU Map: maximal clique based, watershed based method
13-20 Hz Dataset max. clique based water- shed based For higher frequency band (13-20Hz) even more similarity between MCB and WB FU maps.
27
Interpretation Results
Simultaneous activity - at a more global scale for a lower EEG frequency - at a more local scale for a higher EEG frequency [Nunez et al, Electroenceph Clin Neurophysiol 103, 5 (1997)] In accordance, we observe less inter-FU connections for higher frequencies Connections between anterior and posterior FUs are probably associated with the two most important sources of activity for this type of data [Comerchero & Polich, Clin Neurophysiol 110, 1 (1999)] Correspondingly, we observe connections between anterior and posterior FUs This slide is not required for the ISMM presentation. However, it motivates the use of the FU map in general. Given enough time, I would present this slide.
28
Performance maximal clique based and watershed based method
Time complexity Avg. FU detection time Maximal clique based O(3n/3) 24 s Watershed O(n2 log n) 0.06 s For the datasets used, there was hardly any difference in FU detection between datasets. (In our paper submitted to IEEE TVCG, the difference is less consistent but the WB method is always (up to 100,000 times) faster.) n = no. electrodes (up to 512) Average speedup: 400 x
29
Conclusion & Discussion
Functional unit (FU) map FU = data-driven ROI = spatially connected clique data-driven EEG coherence visualization FU detection - Existing method Maximal clique based (MCB) Time complexity O(3n/3) - Novel method Watershed based (WB) Time complexity O(n2 log n) Result: - WB method on average 400x faster than MCB method - FU maps: MCB and WB similar, but Oversegmentation: adapting significance threshold increases similarity future: merge spatially neighboring FUs if together one FU
30
Future Work Case studies: aging, mental fatigue Group analysis
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.