Multinomial Classification Deep Learning Approach for Knee Osteoporosis

Jothsna Praveena Pendyala

Abstract: Early osteoporosis detection is considered crucial for minimizing health consequences. The primary method for early osteoporosis detection is through bone density testing. However, this is time-consuming and not affordable for everyone. As a result, machine learning techniques have been introduced to improve the cost and performance of this early detection process. Unlike previous research, which deals only with osteoporosis detection as a binary classification problem, this research deals with the case as a multinomial classification problem by considering osteopenia as a separate class. This paper studies deep learning-based classification of knee osteoporosis utilizing transfer learning by fine-tuning the layers. The fine-tuning is carried out except for the normalization layers of the following networks: MobileNetV3-Large, InceptionRestNetV2, VGG19, and DenseNet121. These models are trained on an augmented version of the 1947 knee X-ray dataset. To further contribute to the existing research, the proposed method divides knee X-ray pictures into three classes: osteopenia, osteoporosis, and normal. This multinomial classification perspective makes it easier to initiate early screening and treatment. The collected results compare the computational efficiency of MobileNetV3-Large and the powerful feature extraction capabilities of InceptionResNetV2 and VGG19 by measuring model latency and classification performance. VGG19 demonstrated the best classification accuracy of 93.08%, followed closely by InceptionResNetV2 at 91.79%, DenseNet121 at 90.77% and MobileNettV3-Large at 90.43%. MobileNetV3-Large is better in execution time compared to these other models. A comparison with existing models highlights how the proposed fine-tuned models could offer a fast, scalable, and accurate solution for automated osteoporosis screening, supporting early detection and improved patient outcomes.