Komparasi Supervised dan Self-Supervised Learning untuk Klasifikasi Sel Darah pada Skenario Keterbatasan Data Berlabel

Misella Mambu (Universitas Sam Ratulangi)
Oktavian Abraham Lantang (Universitas Sam Ratulangi)
Salaki Reynaldo Joshua (Universitas Sam Ratulangi)

Riau Jurnal Teknik Informatika
Vol. 5 No. 2 (2026): Juli 2026

The development of blood cell image classification models is often hindered by the scarcity of labeled data. This study evaluates the effectiveness of supervised baseline models, supervised fine-tuning, and self-supervised learning (SSL) based on SimCLR and BYOL using the ResNet-50 and DenseNet-121 architectures. Data-limited simulation was performed using 10% labeled data for the classification stage and 90% unlabeled data for SSL pre-training on 17,092 images from 8 morphological classes. The results show that SSL achieves competitive performance, with the highest accuracy on ResNet-50 at 95.31% (SimCLR) and 94.27% (BYOL), and on DenseNet-121 at 93.75% (SimCLR) and 89.58% (BYOL). The SSL approach also demonstrated robustness against overfitting, good robustness against class imbalance, and Grad-CAM visualizations that align with medical expert assessments on ResNet-50. In conclusion, SSL effectively optimizes unlabeled data to improve the performance of classification models in scenarios with limited data annotation.