Development of Modified Logistic Regression (MLR) Model to Classify Psoriasis and Non-Psoriasis Images
pdf

Keywords

Contours
Gaussian Blur
Canny Edge Detection
the Hough Circle Transform
Modified Logistic Regression

Abstract

Purpose: Skin Psoriasis is one type of chronic skin disease that causes the immune system to cause harm to the human body which results in rapid multiplication of skin cells. This causes inflammation and forms scaly patches of skin, mostly on the scalp, elbows, or knees. The exact cause of this disorder is not known but it is suspected to be a combination of genetics and environmental factors. It is not contagious in nature. The task of identification of this disease is very challenging in nature because it resembles with some other skin diseases. So, proper detection of the disorder is needed for effective treatment. This paper is an approach to suspect and detect psoriasis at an early stage with the help of machine learning approach. Some of the features like shape, colour, presence of scale, border etc. are taken into consideration for detection. The images are collected from the Kaggle database and North Bengal Medical College and Hospital for training and testing the system.

Methods: Image data of psoriasis affected skin and normal skin is collected from Kaggele and North Bengal Medical College and Hospital is collected and stored separately. The collected Image data is labeled for whether it shows psoriasis or not. Images are normalized to maintain consistency in the dataset, which helps the machine learning model converge more efficiently. In case of segmentation, pixel-level annotations may be used. Various feature extraction methods such as shape identification, texture identification, colour identification, location identification and border identification. Initially, the features like shapes, borders and texture are recognized using some image processing techniques. One algorithm is designed here in order to find the largest circle or oval which helps in recognizing the shape. Additional methods to extract features are used such as Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), or color histograms. On the basis of the features patterns are developed which are linked to each image where a unique id is created for the image. An Ensemble approach of SVM-Random Forest is used here which is a classical Machine learning approach.

 

Results: The patterns are vertically split into two subsets sample and target. Sample is a vector consists of the fields specifying the features and target is a binary value which divides the images into two classed Psoriasis and Non-Psoriasis image. A total of 2015 images of Psoriasis and 2000 Non-Psoriasis are collected from kaggle and merged into a single database which consists of a mixture of both class of images. These are further horizontally split into train and test database in the ration of Train: Test =80:20. The accuracy was 98.99%. Further 2000 primary images consisting of a mix of both classes is collected from NBMCH and an accuracy of 96.95% was found.

 

Conclusion: The SVM-Random Forest performed commendably and more options will be used in future like deep learning in order to recognize the severity of the disorder

pdf
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Copyright (c) 2024 African Journal of Biomedical Research