Implementation of Pretrained Models to Classify Osteosarcoma from Histopathological Images

Osteosarcoma is the leading contributory factor to mortality and disability in adolescents, teens, and young adults. Despite recent advances in osteosarcoma diagnostic procedures, its exorbitant costs compel researchers to develop a low-cost diagnostic approach, and histology still remains the preferred method for assessing disease stage. Over the past few decades, deep learning approaches have revealed efficacy for analysing as well as categorising histopathological images. Convolution neural network (CNN), a deep learning technique has been more prominent in the field of image analysis. In this study, the efficiency of transfer learning approaches is explored and assessed. The transfer learning techniques specifically Vgg16, EfficientNetB5, ResNet50, and DenseNet169 pretrained CNNs employed on a public dataset to identify and classify osteosarcoma from histopathology images into viable, non-tumor, and tumor groups. The ImageNet model’s weights are reused in the convolution models, along with transfer learning and a fully connected layer that works with a three-class label. The achieved result outperforms previous models, rendering it suitable for clinical cases and thereby supporting society in the early detection of osteosarcoma, lowering the burden on pathologists.