Klasifikasi Multikelas Citra Chest X-Ray Menggunakan Semi-Supervised SoftMatch pada Label Terbatas

M. Nabil Dawami (Universitas Islam Negeri Sultan Syarif Kasim, Pekanbaru)
Benny Sukma Negara (Universitas Islam Negeri Sultan Syarif Kasim, Pekanbaru)
Muhammad Irsyad (Universitas Islam Negeri Sultan Syarif Kasim, Pekanbaru)
Yusra Yusra (Universitas Islam Negeri Sultan Syarif Kasim, Pekanbaru)
Febi Yanto (Universitas Islam Negeri Sultan Syarif Kasim, Pekanbaru)

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 6 No 12 (2026): May 2026

Deep learning-based chest X-ray (CXR) classification frequently encounters bottlenecks due to the scarcity of labeled medical data and imbalanced class distributions. This study aims to implement a semi-supervised learning (SSL) approach utilizing the SoftMatch algorithm with a DenseNet-121 backbone for the multiclass classification of CXR images (Normal, COVID-19, and Pneumonia) under limited label conditions. SoftMatch is specifically selected for its capability to mitigate the quantity-quality trade-off through an adaptive pseudo-label soft-weighting mechanism. A dataset comprising 5,228 images is allocated via a stratified split into 70% training data, 10% validation data, and 20% testing data. Experiments are conducted across three labeled data proportion scenarios (5%, 10%, and 20%), each evaluated with and without Uniform Alignment. Evaluation metrics include accuracy, macro F1-score, confusion matrix, ROC-AUC, supported by visual interpretability analysis using Grad-CAM. The experimental results demonstrate that the model remains robust under the most critical scenario (5% labels), achieving an accuracy of 91.68% and a macro F1-score of 91.72% when integrating Uniform Alignment (UA), outperforming the scenario without UA, which records an accuracy of 90.73% and a macro F1-score of 90.82%. The best performance for the UA configuration is achieved in the 10% label scenario (accuracy 94.46%; macro F1-score 94.58%), while the peak overall performance is attained by the 20% label scenario without UA (accuracy 95.79%; macro F1-score 95.89%). These findings indicate that Uniform Alignment is effective in low-to-medium label conditions but does not consistently enhance performance at higher label proportions.