# Practice Exam Questions

## Practice Exam

The practice exam is not a real exam, but covers some (not all) of the concepts and questions that might be covered by the real exam (which is oral). If you have done a project some (0-50%) of the images or histograms will come from your project slides instead of random samples as it is now. The exam will generally not be multiple choice!

## Choosing a threshold

Based just on the distribution above which value would make the most sense for the threshold

1. 0.5
2. 1.0
3. 1.75

The value 1.0 lies best between the two peaks of the distribution

## Image Enhancement Speckled

Given the following measured image of a sample of a cross, which filter would you expect to work best?

1. Median
2. Gaussian / Low Pass
3. Edge Detection / High Pass
4. Diffusion
5. Other (Explain)

Think about what separates the image from the noise. Are they independent? Are the surrounding pixels a good approximation for the image?

A median filter is the best choice since the noise is randomly distributed and strong in intensity (easily distinguishable from the image). Using the other filters would spread or blur it out (since they are linear) when it needs to be removed from the data and that requires a non-linear filter.

## Image Enhancement Blurry

Given the following measured image of a sample of a cross, which filter would you expect to work best?

1. Median
2. Gaussian / Low Pass
3. Edge Detection / High Pass
4. Diffusion
5. Other (Explain)

Think about what separates the image from the noise. Are they independent? Are the surrounding pixels a good approximation for the image?

A high pass (or edge-detecting, also called sharpening) filter would be best suited for this image since there is no random or speckle noise, instead just a blurred out version of the image (as if the microscope were not focused correctly). Since we see that the blurring only occurs in the x direction, the filter could be further optimized by using a kernel / neighborhood that took this into account rather than performing edge detection in all directions.

## Image Enhancement Background

Given the following measured image of a sample of a cross, which filter would you expect to work best?

1. Median
2. Gaussian / Low Pass
3. Edge Detection / High Pass
4. Diffusion
5. Other (Explain)

Think about what separates the image from the noise. Are they independent? Are the surrounding pixels a good approximation for the image?

A low pass (or specifically Gaussian) filter would be best for these images since the noise is evenly distributed and not limited to single bad pixels. The underlying assumption allowing us to use this filter is that the neighborhood collectively proivdes a estimate of the value than any single pixel.

## Image Enhancement Challenging Question

Given the following measured image of a sample of a cross, which filter would you expect to work best?

1. Median
2. Gaussian / Low Pass
3. Edge Detection / High Pass
4. Diffusion
5. Other (Explain)

Think about what separates the image from the noise. Are they independent? Are the surrounding pixels a good approximation for the image?

The best option for this image is other, since non of the listed filters account for the type of artifcats present in the image. The best filter would be a cut-off which ignored all values above 1.5 since they are all noise (and the variation in the signal portion is very small). While a median filter would work well, it is very general and might distort some of the smaller features. A low-pass would blur out the spikes over many pixels making them even more difficult to remove.

## Morphological Operations

Looking at the before and after images, which morphological operation was most-likely performed and what was the range

1. Erosion, Small Window
2. Dilation, Small Window
3. Opening, Small Window
4. Closing, Small Window
5. Erosion, Large Window
6. Dilation, Large Window
7. Opening, Large Window
8. Closing, Large Window

Pay attention to the edges of the image and the size of the features inside

### Kernel Shape

The operation can be seen by looking at the inside and edges. Since the inside is filled in, it must be either dilation or closing (the other operations typically only remove pixels). By examining the edges and seeing that the structure does not grow we can determine it is closing (dilation would have caused the borders to grow).

### Neighborhood

The size or neighborhood can be guessed by looking at the size of the holes filled which are clearly much larger than 3 pixels.

## Segmentation Strategy

You measure the same cross 4 times and and get the following images as output. Estimate the signal to noise in the images and provide a strategy for how the data could be filtered and segmented.

## Automatic Thresholds

Which is the best reason to use automatic threshold techniques on data

1. Reduce noise
2. Compensate for changing illumination
3. Improve signal to noise ratio
4. Segment difficult to separate phases

Check the "Where segmentation fails" slide

Changing illumination will change the brightness of the pixels in the image but should not largely change their statistics or distribution making them ideal candidates for automated methods.

Neither noise nor signal to noise ratio can be improved using automated segmentation. These are addressed soley in the "Image Enhancement" lecture of the course. While automatic techniques might make results more reliable in 4, difficult segmentations are just as difficult when using automated techniques

## Selecting a threshold technique

Based soley on the histogram above which automatic threshold technique is best suited?

1. Intermodes
2. Hysteresis Threshold
3. K-Means

Intermodes works best since the two phases are well characterized wiht the modes of the system and since their does not appear to be any bias or skew the method taking the value between them would work best.

K-Means and Hysteresis might work as well but it is hard to know without seeing the image itself

## Applying Thresholds

Can a threshold be applied to this image? Why or why not?

## K-Means Clustering

Which statement best describes the limitations and shortcomings of using K-means clustering

1. Since it is an iterative method it can take a long time to converge
2. The random seed start locations can influence the final result and running the analysis multiple times often results in multiple different answers
3. K-Means can only handle position information
4. K-Means can only handle small imaging datasets
5. K-Means is complicated and therefore not many programs support it.

Check the "K-Means" slides in the Advanced Segmentation presentation

The start locations for the clusters in K-means are random and in many situations the start locations greatly influence the final groups (local-minima problem). The issue can be alleviated by running K-means many times on the same dataset but when different data is used it is always possible that the resulting classification is very different. The best way around this is to use K-means to establish a standard set of segmentation or thresholding criteria and then apply these criteria (which are now fixed and not subject to random cluster locations) to the new datasets rather than K-means itself.

While K-means is iterative, it has been optimized and can run very quickly even on massive datasets. K-means has been implemented on systems where handle enormous datasets (much larger than ImageJ or Matlab) and given its vector space formulation can handle nearly any kind of information associated to each point. Every major math or programming tool has support for K-means (Matlab, Java, Python, Octave, R, Paraview, even Excel with an Add-In) and it is very easy to implement.

## Shape Analysis Anisotropy

Explain what anisotropy means in reference to shape and explain why it might be useful when examining cells which are thought to be responsible for mechanosensation in bone?

### Hint

The defitions of anisotropy is in the Analyzing Single Objects lecture and the mechanosensation requires you to think about how forces might interact with such a cell.

## Shape Analysis Validation

Below is a segmented image of cells using two different thresholds. Each image is right in some ways (accurate cell shape and volume) and wrong in some ways (too many noise spots, etc)

How would you choose the best image and validate your method? What additional steps could be taken to further improve the segmented image?

## Distribution Analysis

Below is a plot of object center of volume coming from the result of segmentation and shape analysis. Is there any emergent pattern in the positioning of these cells? Are the positions randomly distributed? Which metrics might be useful for further characterizing the randomness?

## Distribution Analysis

Below is a plot of object center of volume coming from the result of segmentation and shape analysis. Is there any emergent pattern in the positioning of these cells? Are the positions randomly distributed? Which metrics might be useful for further characterizing the randomness?