Pengembangan Model Deep Learning Menggunakan Fine-Tuning ResNet-18 untuk Diagnosis Leukemia Akut pada Blood Smear Images

Abstract

This study presents an investigation of fine-tuning ResNet-18 model for the precise diagnosis of acute leukemia from blood smear images. Early detection of acute leukemia is crucial for improving patient prognosis. Despite advancements in deep learning for image recognition, the utilization of ResNet-18 for acute leukemia diagnosis from blood smear images remains limited. Hence, this research proposed fine-tuning ResNet-18 to enhance accuracy in acute leukemia diagnosis. Two blood smear image datasets, Dataset RS from RSUP Haji Adam Malik Medan and ALLIDB1 from the Università degli Studi di Milano Statale were collected. After image preprocessing, the model underwent training using fine-tuning techniques on the ResNet-18 architecture. Evaluation results demonstrate high accuracy, with 99.12% accuracy on the validation dataset and 99.12% on the test dataset. Additional evaluation metrics, including precision, recall, F1-score, and AUCROC, also exhibit excellent performance in classifying blood smear images as acute leukemia or normal. Comparative analysis with three other architectures, namely ResNet-18 without fine-tuning, VGG-16, and MobileNet V2, reveals that finetuning ResNet-18 yields superior performance in terms of accuracy and stability. This study emphasizes the significance of fine-tuning in enhancing the quality and reliability of models for acute leukemia diagnosis.