ROBUSTNESS OF CONVOLUTIONAL NEURAL NETWORK ON VARIOUS WINDOWING FOR RECOGNITION OF PATTERNS IN CHEST CT IMAGES
Convolutional Neural Network (CNN) has been widely applied to analyze images as diagnostic support. Nowadays, CT Scan is the most commonly used modality for clinical diagnosis. Windowing techniques can be used to interpret the CT images displayed on a computer screen. The function of this techniq...
Saved in:
| Main Author: | |
|---|---|
| Format: | Theses |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/71840 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | Convolutional Neural Network (CNN) has been widely applied to analyze images
as diagnostic support. Nowadays, CT Scan is the most commonly used modality for
clinical diagnosis. Windowing techniques can be used to interpret the CT images
displayed on a computer screen. The function of this technique is to shift focus the
Hounsfield unit (HU) values on the region of interest by adjusting the Window
Width (WW) and Window Level (WL). This research was conducted to find out the
robustness of the CNN model in pattern recognition of chest CT images by varying
the WW and WL values. There are two types of the dataset: CT images of Covid-19
with a total of 589 of 190 patients and non-Covid-19 CT images with a total of 321
of 76 patients. Furthermore, pre-processing of the collected image dataset was
carried out by varying the WW and WL values so that 8,190 images were obtained.
These images were divided into training and testing data in a ratio of 70:30. The
CNN model consists of fixed and variable parameters. The fixed parameters consist
of an image size of 512x512, a learning rate of 0,001, epochs of 150, and the number
of convolutional filters of 32, 64, and 128. Meanwhile, the variable parameters are
as follows: kernel size of 3x3 and 5x5, hidden layers of 1HL and 3HL, the number
of neurons in hidden layers of 32, 64, 128; and weight dropout 20% and 50%. Eight
scenarios were generated from these two parameters and used to evaluate the
model's accuracy, precision, and fluctuation in order to assess CNN's level of
robustness. The results of this research shows that there are fluctuations in the
accuracy and precision of the data training that are quite insignificant at epochs 1
to 4 in model scenarios 1, 2, and 4 for some WW and WL values. This is due to
using optimal weight distribution. The results of testing model scenarios 1 to 8 for
all WW and WL values resulted model accuracy values between 99,65% and
99,98%% and model precision values between 99,73% and 99,98%. Based on the
results of this research, it can be concluded that the accuracy and precision values
for all variations of WW and WL values are always stable in the range of 99,65%
to 99,98%. It shows that the CNN models applied to chest CT images pattern
recognition have a high level of robustness.
|
|---|