<![CDATA[The development of deep learning-based Computer-Aided Diagnosis (CAD) for tuberculosis (TB) detection faces a fundamental challenge: high reliance on massive annotated data, which requires scarce radiological expertise, considerable time, and high costs. This study proposes a Self-Supervised Learning (SSL) approach via the SimCLR framework as a strategy to reduce labeled data requirements in X-Ray-based TB classification. The model utilizes a ResNet-50 encoder trained contrastively on unlabeled data using the NT-Xent Loss, followed by downstream adaptation via linear probing (SimCLR-LP) and fine-tuning (SimCLR-FT). Utilizing datasets from UPF BBKPM Makassar, evaluations were conducted across four labeled data fractions (10%, 25%, 50%, 100%). Results demonstrated that at the 10% fraction, SimCLR-LP achieved 85.50% accuracy and an AUC of 0.9091, significantly outperforming the Baseline model (62.60% accuracy) which suffered from degenerate prediction. The SimCLR-LP variant achieved ≥80% accuracy using only 60 labeled images, whereas the Baseline required 303 images to reach a comparable threshold, demonstrating a fivefold labeling efficiency. Grad-CAM analysis confirmed that SimCLR-FT yielded localized activations in the perihilar and lower lung lobes, unlike the Baseline's scattered activations without anatomical focus.]]>
Copyrights © 2026
