Malaria-Image-Classification

This repository contains code for end-term project of class Digital Image Processing & Applications taught by Prof Deboot Sheet.

Malaria-Image-Classification

The below method is successfull in classifying Malarial/Normal samples with with ~92% accuracy.

For MATLAB Central File Exchange please visit

For GitHub Repo please visit GitHub

Visualize dataset

1. Dataset can be downloaded from Malarial Cell Images Dataset.

2. Dataset consisits of 13780 parasited and 13780 non-infected images. A total of 27560 images are available in the dataset.

Some examples of the dataset are shown below.

Image class	Image 1	Image 2	Image 3	Image 4	Image 5
Parasite Image
Non Infected Image

3. Purple regions can be easily observed in the parasite images which are not available in Non-Infected images.

Classification task

Now the task is to classify the given input image into Malaria or Normal class.

Methods

Gradient-based edge dectection and morphological operations for Malaria image classification

Steps	Description	Result
1	Convert RGB to Gray Image
2	bw = Edge detection using sobel filter
3	bw005 = Edge detection using sobel filter with thresold = 0.05
4	Subtract bw from bw005
4	Dilate the bw005-bw image with structuring element of disk with radius=5
5	If the area of the dilated image is more than the threshold then, image is classified as malarial infected sample. Thresold = 100	Area = 937, So classified as Malaria .

Results

Below are results are obtained by performing classification on 13780 Malaria adn 13780 Normal class images.

Metric	Accuracy	Sensitivity	Specificity	Precision	Recall	F1-Score
Metric Value	91.97	0.929	0.910	0.912	0.929	0.920

Prediction Results

Testing on Normal-class images

Testing on Malaria-class images

Conclusions

The proposed method solves the malaria image classification problem with good accuracy. A lot of other advanced methods like use of Classifiers on top extracted feature using feature extraction and more advanced deep learning algorithms like Deep CNNs. But there is a trade-off between computaional effiecieny, time required to train & inferece. The proposed method solves the problem with least possible resources and time constraints with comparable accuracy to more sophisticated methods.