usage:division_detection

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision		 Previous revision
usage:division_detection [2019/11/15 18:31] – [Validation] pseudomoanerusage:division_detection [2023/07/05 15:36] (current) – [The Division Detection module] pseudomoaner
Line 10: Line 10:
  
 Despite these differences, most of the machinery responsible for performing these assignments remains the same. The layout and usage of the division detection module is therefore very similar to that of the tracking module. For the purposes of conciseness, the below usage guide focuses on the ways in which the two modules differ, rather than providing a complete guide on each aspect of the division detection process. To initialise it, simply press the 'Divisions' button on the home panel. The following GUI should now appear: Despite these differences, most of the machinery responsible for performing these assignments remains the same. The layout and usage of the division detection module is therefore very similar to that of the tracking module. For the purposes of conciseness, the below usage guide focuses on the ways in which the two modules differ, rather than providing a complete guide on each aspect of the division detection process. To initialise it, simply press the 'Divisions' button on the home panel. The following GUI should now appear:
 +
 +{{ :usage:divisiondetectionv2.png?nolink&800 |}}
  
 ===== Model training ===== ===== Model training =====
  
 Training the division detection module is very similar to training the tracking module. The processes of feature choice and training link inclusion proportion remain the same as before, although the relatively low number of division events compared to object-object links means that the histogram used to inform the choice of the training link inclusion proportion may not be very informative. One major difference does exist between the modules however: because temporal information is included as a feature, all 'training' divisions are effectively pooled into a single dataset, rather than being split by timepoint. As a result, the division detection module returns a single trackability score $R$, rather than a separate score for each timepoint. This is indicated in the top right-hand corner of the unnormalized step size distribution: Training the division detection module is very similar to training the tracking module. The processes of feature choice and training link inclusion proportion remain the same as before, although the relatively low number of division events compared to object-object links means that the histogram used to inform the choice of the training link inclusion proportion may not be very informative. One major difference does exist between the modules however: because temporal information is included as a feature, all 'training' divisions are effectively pooled into a single dataset, rather than being split by timepoint. As a result, the division detection module returns a single trackability score $R$, rather than a separate score for each timepoint. This is indicated in the top right-hand corner of the unnormalized step size distribution:
 +
 +{{ :usage:histogram.png?nolink&400 |}}
  
 <note tip> <note tip>
 Because there are no other internal values to compare it to, $R$ is not very useful for determining the ease of division assignment. Instead, it is recommended that the accuracy of division detection be assessed following its completion using the **View divisions** panel and the [[usage:overlays|overlays]] module. Because there are no other internal values to compare it to, $R$ is not very useful for determining the ease of division assignment. Instead, it is recommended that the accuracy of division detection be assessed following its completion using the **View divisions** panel and the [[usage:overlays|overlays]] module.
 +</note>
 +
 +<note tip>
 +Division events are much sparser than ordinary frame-frame links between objects, and typically need less feature information for assignment. Prediction of the location of daughter cells in feature space is also more noisy than the prediction of a single object's location in the subsequent frame, meaning some highly variable features can actually detract from the accuracy of division detection if included. It is therefore recommended that you initially try training the division detection model with only the **Position** feature active.
 </note> </note>
  
 Model training is initialised as before, by clicking the **Calculate!** button. Once it has completed, division detection proper becomes available. Model training is initialised as before, by clicking the **Calculate!** button. Once it has completed, division detection proper becomes available.
- 
- 
- 
  
 ===== Division detection ===== ===== Division detection =====
  
-Selection of the **division threshold** is also very similar to selection of the **adaptive linking threshold** in the tracking module, with the user repeatedly performing a cycle of threshold selection - division detection - validation - threshold update. However, unlike the tracking module, the relative sparseness of division events means that it is practical to perform threshold selection by repeatedly analysing the entire dataset, rather than a single pair of frames. +In contrast to the Tracking module, the Division Detection module does not require the user to manually select a tracking threshold. Instead, the division detection algorithm automates selection of this threshold by iteratively modifying its value such that the resulting lineage structure is maximally connected while still remaining a tree (i.e. contains no loops). In practice, this generally gives a near-optimal balance between assigning true links - correct assignments between mother and daughter cells - and suppressing false links.
- +
-To perform initial division detection, simply click the **Find divisions!** button. One of two outcomes will then occur: +
- +
-  - If division detection completed successfully, the division viewer and lineage distribution windows will update. +
-  - If division detection resulted in a lineage with a cycle (i.e. a cell marked as its own ancestor), the following warning notice will appear and the division detection process will be aborted: +
- +
-In the case of scenario (2), the **division threshold** should be reduced until the warning notice ceases to appear. +
- +
  
 +Once training has completed, it is therefore sufficient to simply press the **Find divisions!** button to launch this iterative algorithm and generate your lineage.
 ==== Validation ==== ==== Validation ====
  
Line 47: Line 44:
   * The division viewer (bottom left-hand corner) acts similarly to the test tracking viewport. For each division event, maternal cells will be marked in yellow, while the corresponding daughter cells will be marked in purple. To switch between mother and daughter timepoints, click the **Toggle A/B** button in the **View divisions** panel. To move to the next division event in the series, click the **Next division** button.   * The division viewer (bottom left-hand corner) acts similarly to the test tracking viewport. For each division event, maternal cells will be marked in yellow, while the corresponding daughter cells will be marked in purple. To switch between mother and daughter timepoints, click the **Toggle A/B** button in the **View divisions** panel. To move to the next division event in the series, click the **Next division** button.
  
-{{ :usage:abtoggle.png?nolink&400 |}}+{{ :usage:divabtoggle.png?nolink&600 |}}
  
   * The lineage size distribution is similar to the track length distribution. It indicates the total number of tracks incorporated into each cell lineage. For example, a cell that divided 3 times would result in a lineage that contained 15 tracks - this cell would be indicated in the lineage size distribution as an additional count in the bin at 15. The **Minimum lineage size** variable can be used to define a minimum cut-off in this lineage size distribution, allowing lineages that are below this threshold to be excluded from the final output. This is indicated in the distribution as a vertical red line.   * The lineage size distribution is similar to the track length distribution. It indicates the total number of tracks incorporated into each cell lineage. For example, a cell that divided 3 times would result in a lineage that contained 15 tracks - this cell would be indicated in the lineage size distribution as an additional count in the bin at 15. The **Minimum lineage size** variable can be used to define a minimum cut-off in this lineage size distribution, allowing lineages that are below this threshold to be excluded from the final output. This is indicated in the distribution as a vertical red line.
Line 65: Line 62:
  
 ===== Video demonstration ===== ===== Video demonstration =====
 +
 +{{ youtube>e6IF4OKy3v0?large }}
  • usage/division_detection.1573842680.txt.gz
  • Last modified: 2019/11/15 18:31
  • by pseudomoaner