Understanding preprocessing script

[thuhoainguyen]

First we need to understand what the script does. I suggest you read and comment (line by line) the code flow. If you have any trouble please comment here and we can discuss.

[thuhoainguyen]

I don't understand the second argument this function:

def resample_image(image, new_spacing=[1.0, 1.0, 1.0]):
    original_spacing = image.GetSpacing()
    original_size = image.GetSize()

[thuhoainguyen]

Basically, the segmentation_preProc.py file crop the large axial abominal CT image to get the small specific kidney bound image for further processor techniques.

[anhtuduong]

@thuhoainguyen I think your description is ok but I may need more info:

• What is the input?
• What is the output? The output will be stored in a new place?
• What kind of algorithm being used here?
• What is the flow of the process here? (A diagram can help)

Next time when you comment pleases tag me (@anhtuduong)

[anhtuduong]

@thuhoainguyen I made a diagram for the classes and the workflow of the code to help you: https://drive.google.com/file/d/1amNT1wPpbPOwA9weN3dbkXe1QY7vUEhF/view?usp=sharing

[thuhoainguyen]

Thank you very much! I will check tomorrow and let you know.

Buonanotte! hoaithu

On Fri, 24 May 2024 at 00:25 Anh Tu Duong @.***> wrote:

@thuhoainguyen https://github.com/thuhoainguyen I made a diagram for the classes and the workflow of the code to help you:

https://drive.google.com/file/d/1amNT1wPpbPOwA9weN3dbkXe1QY7vUEhF/view?usp=sharing

kits23-preprocessing-class.png (view on web) https://github.com/thuhoainguyen/kits23/assets/76017474/b3648d7d-bdad-49c0-94d4-d5475be01c2e

kits23-preprocessing-workflow.png (view on web) https://github.com/thuhoainguyen/kits23/assets/76017474/5912b7ae-fa27-4095-ad55-71445c6ad39d

— Reply to this email directly, view it on GitHub https://github.com/thuhoainguyen/kits23/issues/1#issuecomment-2128137826, or unsubscribe https://github.com/notifications/unsubscribe-auth/BHR373VFUT5IYBDNC3SNL4DZDZUGDAVCNFSM6AAAAABH3BWR4SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMRYGEZTOOBSGY . You are receiving this because you were mentioned.Message ID: @.***>

[anhtuduong]

The output of the preprocessing pipeline will differ significantly from the original NIfTI (.nii) images in several ways due to the various preprocessing steps. Here's a detailed explanation of the differences:

1. Resampling

Original Image:

• The original NIfTI image might have anisotropic spacing, meaning the distance between pixels can differ in the x, y, and z dimensions.

Resampled Image:

• The resampled image will have uniform spacing (1.0 mm x 1.0 mm x 1.0 mm) in all three dimensions.
• This standardization facilitates easier and more accurate comparisons between different images and ensures consistent input for machine learning models.

2. Normalization

Original Image:

• The original image might have a wide range of intensity values, which can vary significantly between different scans and different machines.
• Intensity values might not be normalized, making it challenging to use these images directly for machine learning models.

Normalized Image:

• The normalized image will have intensity values mapped to a [0, 1] range.
• The normalization process uses Gaussian distribution parameters derived from the histogram data, ensuring that the intensity values are adjusted according to the expected distribution of tissue types.
• This helps in reducing the variability due to different imaging conditions and scanners.

3. Cropping

Original Image:

• The original image includes the entire scan volume, which might contain a lot of background and irrelevant information.

Cropped Image:

• The cropped image focuses on the region of interest (ROI) around the kidney, tumor, and cyst.
• It includes a margin of 63 pixels around the segmented region, ensuring that relevant context is preserved while excluding unnecessary background.
• This reduces the size of the image and focuses computational resources on the most relevant areas.

---

Histogram in Preprocessing:

In your preprocessing code, the histogram plays a crucial role in normalizing the intensity values of the images. A histogram is a graphical representation of the distribution of intensity values in an image. It consists of bins, where each bin represents a range of intensity values, and the height of each bin represents the number of pixels (or voxels) that fall within that range.

The histograms for different labels are loaded from a precomputed file. This data likely represents the distribution of intensity values for different tissue types (e.g., kidney, tumor, cyst) across the training dataset.

Computing Gaussian Fits:

• This function estimates Gaussian distribution parameters (mean and standard deviation) for each label from the histograms.
• bin_centers are calculated to represent the midpoint of each bin.
• For each label (e.g., kidney, tumor), the mean and variance of the intensity values are computed using the histogram data.

Range Normalization Using Histogram Data:

• This function normalizes the intensity values of the resampled image using the Gaussian parameters computed earlier.
• For each label of interest (1: kidney, 2: tumor, 3: cyst): -- The quantiles of the Gaussian distribution are computed. -- The image is clipped to the range defined by the extreme quantiles to avoid outliers. -- The intensity values are interpolated to a [0, 1] range based on these quantiles.
• The resulting normalized_image has standardized intensity values, making the data more consistent for further analysis.

Benefits of Using Histogram in Preprocessing:

Standardization:

• Normalizing intensity values across different images ensures that variations due to different scanning parameters or patient characteristics are minimized.
• This helps in making the images more comparable and improves the performance of machine learning models trained on these images.

Contrast Enhancement:

• By focusing on the quantiles derived from the Gaussian fits, the normalization process enhances the contrast between different tissue types, making it easier to distinguish between them.

Outlier Handling:

• Clipping the intensity values to the range defined by the extreme quantiles helps in reducing the impact of extreme values or artifacts in the images.

Summary: In summary, the histogram data in your preprocessing code is used to estimate the distribution of intensity values for different tissue types. These distributions are then used to normalize the intensity values of the images, ensuring that they fall within a standardized range. This process enhances the contrast and consistency of the images, making them more suitable for subsequent analysis and machine learning tasks.