Modified Resnet-50 Architecture for Scoliosis Detection

Scoliosis is a condition where the spine curves abnormally, which can cause discomfort, pain, and difficulties with movement. It is essential to detect and diagnose scoliosis as early as possible to prevent further complications and improve treatment outcomes (Brackett, 2023). The main goal of this study was to improve classification accuracy of ResNet-50 architecture in detecting scoliosis on unclothed human back images, enabling early detection and intervention to prevent the progress of the spine curvature. The modified ResNet-50 architecture in this study incorporates global average pooling and reduces the size of the fully connected layers in the original ResNet-50 architecture. The data used in this study consists of images of normal and with scoliosis unclothed human back images. The dataset was sourced from public repositories, private individuals and patients at President Ramon Magsaysay Memorial Hospital Iba, Zambales. These images were annotated and validated by medical experts from PRMMH. The Modified ResNet-50 model showed outstanding performance with slight fluctuation in validation loss similar to the findings in the study of Artates et. al (2024) that despite of minimal validation loss fluctuations the model can still be more robust and reliable. The Modified ResNet-50 model achieved impressive results and outperformed the baseline ResNet-50 across multiple evaluation metrics. The Modified ResNet-50 model reached an accuracy of ninety-seven percent (97%), both precision and recall values of ninety-six-point five percent (96.5), and F1-Score, Macro & Weighted Average of ninety-seven percent (97%). These results indicate that the model is highly effective in accurately classifying unclothed human back images.