Knee osteoarthritis (KOA) is a prevalent and debilitating joint disorder demanding accurate and timely diagnosis for effective management. Magnetic Resonance Imaging (MRI) offers a powerful tool for KOA assessment. Nevertheless, training 3D deep learning models on 3D MRI images tends to be costly and complex, also the use of the whole region of the MRI scans leads to including noisy and irrelevant features in the model training. This study presents a multitask learning approach to enhance KOA detection through the identification of knee structures from 2D MRI slices. We propose a new framework that simultaneously segments relevant knee anatomical structures from MRI masks and detects KOA presence, capitalizing on shared representations learned across these interconnected tasks. Our Multi-task-OA model trained and evaluated on a large dataset composed of 105,919 MRI slices along with their corresponding masks, significantly outperformed traditional single-task and volume-based learning methods, achieving a remarkable 96.70% accuracy in KOA detection.