Age estimation from left-hand radiograph is a commonly used radiological tool for determining unidentified age. Magnetic resonance imaging (MRI) is particularly useful in age estimation. The wrist MRI is an effective means of evaluating age in adolescents. inadequacy of properly labelled dataset inhibits AI systems from achieving equivalent accuracy to experts. Besides, it takes a considerable amount of time and effort to generate accurate labels and find sufficient dataset There were attempts to provide computer-assisted and Machine Learning based assessment for determinations of bone age. The application of Deep Learning in age estimation, however, is often performed for pediatric purposes and most studies are focused on supervised learning. Therefore, the current study proposes a semi-supervised classification model for estimating adult Chronological Age from wrist MRI. The goal is to leverage a large amount of unlabeled data to improve an accuracy of the proposed model. In turn, making the subjectivity of the assessment consistent during interpretation. Total of 3801 male left hand wrist MRI dataset and 357 reports collected from St. Paul Millennium Medical College Hospital MRI centre Addis Ababa Ethiopia. An adequate set of a report for labelling was not available and requires tremendous effort and time. Stratified 10-fold cross-validation was used for splitting the data into training and validation set. Transfer learning and data augmentation techniques applied jointly to mitigate small dataset and while combating overfitting in parallel. The automatic image annotation tool is developed to annotate unlabeled MR images from model prediction. The semi-supervised model is trained enhancing the image using image preprocessing, which resulted in a loss of 0.1 and mAP of 74.29%. The proposed model achieved an average accuracy of 94.44%. The study could contribute to the medical profession by providing a system that supports experts to estimate chronological age. In this study, automatic image annotation improved model performance through iterative training. However, its accuracy depends on model prediction accuracy. Even though transfer learning from different domain improved the model performance, it wasn't quite as anticipated. Thus, researchers involved in the medical application should make their dataset and pre-trained model open to the public.