Development and research of a modified convolutional neural network for malaria cell pattern recognition
A review and analysis of known solutions of the problem of detecting malaria from images of patients' blood at the cellular level using various machine learning algorithms, including the support vector method, deep belief network, and convolutional neural networks, was conducted. Models based o...
Збережено в:
| Видавець: | Інститут проблем реєстрації інформації НАН України |
|---|---|
| Дата: | 2023 |
| Автори: | , , , , |
| Формат: | Стаття |
| Мова: | Ukrainian |
| Опубліковано: |
Інститут проблем реєстрації інформації НАН України
2023
|
| Теми: | |
| Онлайн доступ: | http://drsp.ipri.kiev.ua/article/view/287018 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
Організація
Data Recording, Storage & Processing| Резюме: | A review and analysis of known solutions of the problem of detecting malaria from images of patients' blood at the cellular level using various machine learning algorithms, including the support vector method, deep belief network, and convolutional neural networks, was conducted. Models based on neural networks demonstrate greater efficiency. In particular, all models based on Deep belief network and convolutional neural networks show a classification accuracy of more than 95 %.
It was decided to develop our own model based on a convolutional neural network, which turned out to be the most promising algorithm among those considered. In the development of the proposed solution, a publicly available set of annotated images of patient blood cells was used, which was corrected according to other work that considered this data set. The Python programming language was used in combination with the TensorFlow library, which was applied directly to develop the network. The OpenCV on Wheels library was utilized to resize images from the dataset.
The model consists of 16 layers: 5 convolutional, 5 aggregating, one dropout layer and 5 fully connected. After the development of the machine learning model, the accuracy of the model was tested and compared with the analogues discussed above. Testing was performed independently on two data sets: a set consisting of images scaled to a size of 50×50 pixels and a set consisting of images scaled to a size of 100×100 pixels. According to the test results, it was established that the model is at the level of the best considered analogs based on convolutional neural networks in terms of classification accuracy of test data, having a classification accuracy of 96,68 % and 98,08 % on a set with smaller and a set with larger images, respectively. The model reaches these values at about the fifteenth epoch of training, and the phenomenon of overtraining is observed in the following epochs. |
|---|