Scribd
Upload
61 views

ECG Digitization

The document discusses digitizing electrocardiogram (ECG) scans to make them understandable to AI systems. It describes pre-processing steps to: [1] Remove noise from scanned ECG images; [2] Extract individual curves; and [3] Convert the curves to grayscale data. This involves techniques like thresholding, morphological operations, component labeling and tracking algorithms. The goal is to generate clean grayscale ECG data from raw scans to allow AI models to analyze heart conditions.

Uploaded by

Sushil Acharya
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
61 views

ECG Digitization

The document discusses digitizing electrocardiogram (ECG) scans to make them understandable to AI systems. It describes pre-processing steps to: [1] Remove noise from scanned ECG images; [2] Extract individual curves; and [3] Convert the curves to grayscale data. This involves techniques like thresholding, morphological operations, component labeling and tracking algorithms. The goal is to generate clean grayscale ECG data from raw scans to allow AI models to analyze heart conditions.

Uploaded by

Sushil Acharya
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 23

ECG Digitization

Mao, Junyan
Problem

An Electrocardiogram (ECG) keeps track of the electrical signals in your heart. It’s a common
test used to detect heart problems and monitor the heart’s status in many situations. Currently,
a lot of ECG records, unfortunately, are only stored in printed form, which cannot be processed
or analyzed by a computer algorithm or an AI. While a human doctor would be able to make
initial diagnoses on an ECG just based on bare sight, such a ‘noisy’ scanned ECG (example on
next slide) would make no sense and is unlearnable to an AI. This project aims at developing a
Python-based pre-process procedure, with computer vision techniques, that makes the original
ECG and internal cavity ECG more understandable and learnable by an AI.
Example of ECG and internal cavity ECG
(respectively)

noisy clean
Problems with scanned ECG

1. The dots in the background are bare noises, which prevent the AI
from learning anything, and need to be eliminated.
2. The outside boundary does not help with our analysis, it needs to
be cropped out.
3. There are some gaps in a single curve, which will cause the curve
to be recognized as 2 parts instead of 1 continuous curve, so they
need to be filled.
4. The labels (I, II, V1, V2, …) for the curves also need to be removed
before we extract the curves.
Problem: Dots in the background

● Read the image as grayscale


○ Since color dimension is not needed
● Apply Gaussian Blurring and Median Blurring to the image
○ To reduce image noise and details
● Apply Adaptive Thresholding to the image
○ To further reduce background noise and binarize the image

Result on next slide


Problem: Dots in the background

● We can tell that the background dots are


almost all eliminated while preserving the
shapes of the curves.
● There are still several white dots in the
background, we will deal with that in a later
step
Problem: Outside boundary

● Apply an masking algorithm that approximates the outer boundary, cropping out the
empty areas.
● Further crop a few pixels to totally eliminate the boundary.

Result on next slide


Problem: Outside boundary
Approximating boundary Further cropping

New problem: the highest part of V1 is actually connected to the boundary.


Problem: Gaps in a single curve

● Apply a labelling algorithm on current binary


image.
● We can tell that there are several curves that are
identified as two or more separate curves, which is
caused by the gap.

Several examples on the right


Problem: Gaps in a single curve

● Fill the gaps by applying dilation + erosion


● Dilation
● Erosion
● Label again
● We can now see that the curves are
separated correctly

New problems:

1. When doing dilation and erosion, the ‘aVR’


label was merged into that curve.
2. Some of the peaks are filled due to this
process.
Problem: Extracting individual curves

● Apply another labelling algorithm in SciPy’s


ndimage package to get access to each
connected component
Problem: Extracting individual curves

● Apply another labelling algorithm in SciPy’s


ndimage package to get access to each
connected component
● Filter out small pieces and noises by inspecting
the shape of the components. We can now see
that 12 individual curves have been extracted.

Note: You might see in some examples, there are


parts of another curve being shown. This does not
mean that they are of the same component. It is
because when we are cropping the curve from the
original image, we crop the whole area that a curve
occupies, which might include some parts of another
curve
Now that the individual curves have been extracted. We can continue to do signal extraction and
grayscale conversion. This will be introduced later on the example of internal cavity ECG in this
presentation.
Internal cavity ECG

● This is an example of an internal cavity


ECG, which are screenshotted from a
computer.
● Since the background is solid black, there is
basically no additional noise, unlike the
scanned ECG.
Step 1: Extracting the curves

● Read the image as grayscale.


● Since we only want to extract the five green curves, we crop out the scale and the white
curve above, and the tag below.
● And we binarize the image to reduce dimension.

Result
Step 1: Extracting the curves

● Label the image.


Step 1: Extracting the curves

● Use component’s shape to filter out small pieces and noises, to get only long curves.
Step 2: Trimming the curves

● Since we want to generate a grayscale image in the future, we want to make sure that the
curves are of same length.
● Iterate through the extracted curves to check the lengths, trim longer curves from tail to
match the shortest curve.
Step 3: Find the baseline for grayscale
conversion

● We want to define a reference axis for each curve so that we can record the amplitude of
each peak. The baselines need to be at the respectively same location for each curve so
that they are on a same scale.
● For human eyes, it’s easy to spot a reference axis just by looking at it. But for a computer
algorithm, we need something more specific and accurate. In this case, we find all the ‘s’
characters in the tag in front of each curve, using filtering and OCR, and define the lower
border of ‘s’ as the reference axis for each curve

Tag ‘S’
Step 3: Find the baseline for grayscale
conversion

● Find all the ‘S’ in the original image,


and find the range of the curves.
● The lower bound of ‘S’ would be the
reference axis for the corresponding
curve.
○ In this case, line 107, 242, 376, 511,
645 are the axes, and they all fall
within the range of the curve.
Step 4: Extract curve data from curve image

● We have the extracted curve from the raw image. We now want to extract the actual data
each curve is containing.
● Scan through each curve, run a tracking algorithm that preserves the shape of the curve.
● We can see that it is almost identical to the original curve image.
Step 5: Generate the grayscale image

● Using the baselines we mentioned before, we are able to generate a grayscale image
which contains the information of the current ECG. Below is a specific area on the
grayscale image.
● The size of the image should be (5, length of curve), since we have 5 curves on a single
ECG combined into 1 grayscale image.

Note: based on the size (width) of our curves, we designated the baseline to be
grayscale == 127, whiter part means that it’s a peak, and darker part means
that it’s a valley.
Future Work

● The precision of curve extraction on internal cavity ECG can still be improved.
● Find a better and less calculation-heavy way to denoise the scanned ECG.
● Find a way to do local dilation and erosion to avoid the situation where the label was
merged with the curve.
● Scale up to a larger image dataset and add ML capabilities to make possible diagnoses on
the generated grayscale image.

You might also like