Detection of musculoskeletal disorders is crucial for the diagnosis and treatment of related conditions, allowing timely interventions and improved patient outcomes. Despite significant advances in imaging technologies, radiologists often face challenges such as a large workload and limited resources. Although convolution neural networks (CNN) have been developed to detect musculoskeletal disorders (MSD), their performance for specific studies, such as finger and hand radiographs, often yields suboptimal Cohen kappa scores. To overcome these limitations, a CNN architecture that employs a triplet network with 128- and 256-dimensional feature embeddings is proposed to enhance classification performance for detecting finger abnormalities in MSD. Preprocessing techniques, including contrast-limited adaptive histogram equalization (CLAHE), are applied to improve the quality of input images. Advanced loss functions, such as triplet loss and focal triplet loss, are used to efficiently optimize feature embedding spaces. Subsequently, a classifier is trained using these feature embeddings to differentiate between normal and abnormal cases of MSD with high precision. This methodology offers a robust and scalable solution to support radiologists by automating abnormality detection, reducing diagnostic delays, and ultimately improving patient care and operational efficiency.