This study presents a novel approach to detecting early-stage Osteoarthritis (OA) in athletes using modified Convolutional Neural Networks (CNNs). The baseline model, EfficientNet B0, has been used in OA diagnosis in the past but faces challenges in accuracy and speed. The researchers addressed these issues by replacing the traditional Squeeze-and-Excite (SE) block with an Efficient Channel Attention (ECA) module; this reduced computational cost and improved efficiency while retaining accuracy. The Knee Osteoarthritis Severity Grading Dataset was used on 1,600 knee X-rays, split between healthy and early-stage OA knees. The images were preprocessed with noise-reduction, normalization, resizing, augmentation, and contrast enhancement. The EfficientNet B0 + ECA model achieved 91.5% accuracy, 89.5% precision, and 92% recall, outperforming larger models while requiring fewer computational resources. Class Activation Maps also showed the model focused on clinically relevant features such as joint gap narrowing and osteophyte formation, indicating that its decisions aligned with detection of early-stage OA. The efficiency and accuracy of the EfficientNet B0 + ECA model supports its use in real-time clinical settings.