Enhanced Residual Osteosarcoma Detection Using Deep Learning And Formularization Techniques On Histopathological Images

Osteosarcoma has become increasingly prevalent globally in recent years, significantly impacting the mortality rates of those affected. Histopathological imaging plays a crucial role in the diagnostic process. The proposed research evaluates the effectiveness of these techniques in accurately detecting necrotic, non-viable, and viable tumors within histopathological tissue. The dataset underwent pretreatment using Gaussian filtering to enhance image quality, followed by formularization techniques to improve model generalization and reduce overfitting. Transfer learning models such as EfficientNetB6, DenseNet201, and MobileNetV2 were employed and trained on histopathological images to improve diagnostic accuracy. Pre-trained models derived from the ImageNet architecture were specifically applied for cancer detection. Additionally, a formularization technique was incorporated into the IGDOOD (Iterative gradient descent Of Osteosarcoma Detection) framework to mitigate overfitting and enhance model performance. Iterative gradient descent was the main optimization algorithm used for training the deep learning model, with the formularization techniques implemented to fine-tune the learning rate improve accuracy, and automatically capture tumor regions to extract the nuclear characteristics of tumor cells. These features to develop a histological image classifier for osteosarcoma, using IGDOOD to predict recurrence and survival rates post-treatment. Performance metrics be approximated to assess the efficiency of osteosarcoma detection with reported accuracy on test data being 95.02 % for EfficientNetB6, 99.10 % for DenseNet201, and 99.40 % for MobileNetV2.