Overview / Usage

The "Deep Learning based Web App for Malaria Parasite Detection in Granular Blood Samples" technology uses a web application to recognize malaria parasites in blood smear images using deep learning techniques. Obtaining and annotating a range of blood smear image collections is the initial stage. Deep learning models for picture categorization, including ResNet50, VGG19, and a modified CNN, are constructed using this dataset. Blood smear images can be submitted for analysis by users of the web program. When uploading images, they undergo preprocessing before being fed into deep learning models. Whether or not parasites have been discovered is disclosed to users. An image depiction of the detection findings shows the zones that have been found highlighted. The solution is set up on a server, its accuracy and other metrics are evaluated, and data security and privacy measures are implemented. Constant model upgrades, user support, and maintenance are needed for this malaria detection system.

Methodology / Approach

Based on the findings of the deep learning-based malaria detection investigation, it is plausible to conclude that deep learning models have showed promise in identifying malaria from tiny blood cell images. This study compares the performance of various deep learning models for recognizing malaria parasites. The results showed that the models have outstanding levels of sensitivity, specificity, and accuracy, indicating that they can reliably diagnose malaria. To make our technique more accessible, we integrated it with a web application. Users can send images of their cells, and the program will display whether or not malaria is present. The study also underlines the importance of having a large and diverse dataset for training deep learning models. Adding more layers to the model and training it on a large number of photographs can both improve performance. It also emphasizes the need for future research to improve deep learning model performance and develop more resilient models capable of handling fluctuations in image data. Overall, deep learning-based malaria detection can provide a reliable and effective diagnostic tool for early malaria diagnosis, hence slowing the disease's spread and improving patient outcomes.

Technologies Used

White blood cells (WBCs) are filtered out of the image before they can be recognized. Following that, the presence of malaria is determined by examining red blood cells. The IGMS technique can be used to find prospective parasite candidates by finding the lowest intensity regions of a grayscale image. A ring-shaped zone with a radius of 22 pixels is generated around each single pixel discovered. D. Feature Extraction: Diagnostic testing can help separate true parasites from background noise. Once the parasite candidates have been extracted, a CNN model is used. The improved CNN model includes softmax, fully connected, max pooling, and CN layers. After each convolutional layer, batch normalization layers are applied, followed by an R.E.L.U activation function. A maxpooling layer is added for every two Cn layers selected. The final feature map for CN is linked to The CNN model contains three separate fully linked layers, each with a unique label. To make the model less complex. This custom-made C.N.N. has several advantages over pre-trained networks, including a lower runtime and input size based on the normal parasite size in thick smear images. The personalized CNN model surpasses pretrained networks in accuracy, while having fewer layers and a shorter runtime. On a common Android smartphone, the system can identify parasites in 10 seconds using an input image measuring 4032 X 3024 X 3 pixels. It is designed as an Android app that connects the smartphone lens to the microscope's eyepiece. After adjusting the microscope to find the correct place in the blood smear, the user can use the app to capture photographs. The OpenCV4 Android SDK library is used to implement all image and video recognition algorithms. The Convolutional Neural Network (C.N.N.) classification model is built using convolution, maxpooling, and batch normalization. The number values above these cuboids indicate the size of the recovered features. Improving the volume and quality of training data, as well as selecting the appropriate architecture, can all aid neural network models in performing better. Feature retrieval within a CNN is possible by integrating one or more convolutional layers into the network. These layers use filters to convolve the input image, resulting in feature maps that highlight key patterns.

Other links

• Paper Publication