GAN-OST: GENERATIVE ADVERSARIAL NETWORKS FOR PRECISION SEGMENTATION AND SYNTHETIC AUGMENTATION OF OSTEOSARCOMA TUMOURS IN MEDICAL IMAGING

Osteosarcoma is an extremely malignant bone cancer that the accurate segmentation of tumour is important for effective diagnosis and treatment planning. One of the primary challenges facing segmentation models is the paucity of human annotated medical imaging data. In this paper, we introduce a generative adversarial network (GAN) -based method for accurate segmentation and synthesis augmentation of osteosarcoma tumours in medical imaging — GAN-OST. The suggested approach utilizes a U-Net styled generator to achieve precise tumour segmentation and PatchGAN-based discriminator for creation of high-quality synthetic images. To make up for the lack of training data, GAN-OST adopts a conditional GAN to create realistic synthetic tumour images that are difficult to distinguish from real tumours and thus can be used as a supplement for model training so as to improve the generalization of the model. We will evaluate the performance of our model with Dice Coefficient, Intersection over Union (IoU), Sensitivity, Structural Similarity Index (SSIM) and Fréchet Inception Distance (FID). We evaluate our model on these metrics to achieve a comprehensive evaluation of segmentation accuracy and image quality. Experimental results on publicly available osteosarcoma datasets demonstrate the superior performance of GAN-OST relative to traditional segmentation approaches, including a remarkable increase in both segmentation precision and generalization. Additionally, the synthetic data created by GAN-OST successfully compensates for the lack of data in small sets and allows reliable acting together data laden proposal. This work demonstrates the advancement of osteosarcoma tumour segmentation and provides a network used for data augmentation that could greatly help other rare cancer types and multimodal imaging scenarios in future research.