|
Pre-trained CNN-based feature extraction for automatic morphological identification of Acanthamoeba spp. |
|---|---|
| รหัสดีโอไอ | |
| Creator | Srisupang Thewsuwan |
| Title | Pre-trained CNN-based feature extraction for automatic morphological identification of Acanthamoeba spp. |
| Contributor | Thitiporn Pramoun, Theekapun Charoenpong, Potchara Tangtragulwong |
| Publisher | Mahasarakham University |
| Publication Year | 2569 |
| Journal Title | Journal of Science and Technology Mahasarakham University |
| Journal Vol. | 45 |
| Journal No. | 3 |
| Page no. | 389-399 |
| Keyword | Acanthamoeba spp., morphological identification, pre-trained CNN, feature extraction |
| URL Website | https://li01.tci-thaijo.org/index.php/scimsujournal |
| Website title | Journal of Science and Technology Mahasarakham University |
| ISSN | 1686-9664 (Print), 2586-9795(Online) |
| Abstract | Acanthamoeba spp. are free-living protozoa and are known to cause severe infections in humans. Traditional morphological identification relies on assessing the size and shape of the inner (endocyst) and outer (ectocyst) walls of cysts, which are categorized into three groups (GI, GII, GIII). However, this method is time-consuming and requires skilled experts. This study aims to develop an automated image analysis system for the classification of Acanthamoeba spp. cysts by employing a pre-trained convolutional neural network (CNN)-based feature extraction approach combined with a Support Vector Machine (SVM) classifier to increase diagnostic accuracy in pathology and reduce classification errors. The feature extraction is performed using various CNN models pre-trained on ImageNet, including Xception, EfficientNet-B0, EfficientNet-B1, VGG16, ResNet50, ResNet101, MobileNet, Inception-V3, and InceptionResNet-V2, to extract high-level global features from microscope images of cysts. These extracted features are then input into an SVM for classifying the cysts into groups GI, GII, and GIII. Additionally, the performance of this approach is compared with that of conventional feature extraction methods. Experimental results demonstrated that the system employing pre-trained CNNs combined with SVM achieved an accuracy exceeding 95%, with ResNet50 and ResNet101 yielding the best results, respectively. These findings validate the potential of using pre-trained CNN-based feature extraction for the robust and efficient classification of complex morphological cyst groups, thereby offering a promising tool for enhancing clinical diagnostic procedures. |
