<![CDATA[Accurate identification of red blood cell (RBC) morphological abnormalities is essential for anemia screening and hematological assessment; however, manual microscopic examination remains time-consuming, subjective, and highly dependent on expert availability. While recent deep learning studies have demonstrated promising accuracy in RBC classification, many focus primarily on model performance without addressing practical deployment constraints or system-level integration for routine laboratory use. In this study, a web-based prototype system for automated RBC abnormality classification is proposed using a lightweight MobileNetV2 architecture. The dataset consisted of 1,320 microscopic blood smear images collected from Klinik & Laboratorium Parahita in Jember and Surabaya, covering six RBC categories with balanced class distribution. All images were anonymized and verified by a certified clinical pathologist prior to use. The model was trained using transfer learning and evaluated on a held-out test set to assess generalization performance. The proposed model achieved a test accuracy of 89.77%, with consistent precision, recall, and F1-score across classes, indicating reliable multi-class classification performance. Analysis of misclassified samples revealed uncertainty primarily between morphologically similar RBC types, reflected by lower confidence scores. These results demonstrate that lightweight deep learning models can provide effective and efficient support for RBC morphology analysis when integrated into an accessible web-based system. The proposed approach contributes a deployment-oriented diagnostic support tool that has the potential to assist laboratory professionals by improving screening efficiency and consistency while preserving clinical oversight.]]>
Copyrights © 2026
