Optical coherence tomography (OCT) scanning is crucial for the diagnosis of widespread ophthalmic diseases. Traditionally, experts manually identify diseases and biomarkers from OCT scans. Recently, modern medical imaging practices have increasingly utilized deep learning techniques to speed up and improve diagnostic accuracy in ophthalmology. However, obtaining accurately labeled datasets is a significant challenge in medical imaging due to the expertise required for precise annotation by trained professionals. This article presents a novel method, named OBoctNet, with a new two-stage training strategy to improve the identification of ophthalmic biomarkers using OCT scans from the OLIVES dataset, which contains only 12% labeled data. This approach leverages a robust methodology that effectively uses labeled and unlabeled data to enhance biomarker identification accuracy, achieving a cumulative performance increase of 23% across 50% of the biomarkers when compared to the previous studies. To better identify biomarkers, the OBoctNet employs an active learning strategy that uses unlabeled data and dynamically ensembles models based on their performance within each experimental setup. Additionally, the usage of Gradient weighted Class Activation Mapping (GradCAM) helps identify regions of interest associated with relevant biomarkers, enhancing interpretability and transparency for potential clinical adoption.