DefinePK

Abstract

Osteosarcoma is an aggressive primary bone malignancy, and the accurate segmentation of technical regions in medical imaging is critical for diagnosing, treatment planning, and prognosis. Yet the lack of annotated data for medical imaging is a key bottleneck for building strong deep learning models. To solve this, we introduce GAN-OSMOSIS, a novel Generative Adversarial Network (GAN)-based framework for enhanced optimized osteosarcoma segmentation and multimetric assessment (MME) via synthetic imaging assist. We used a U-Net-based generator to perform semantic segmentation and a PatchGAN discriminator that classifies images as real or fake based on overlapping patches. Moreover, a Conditional GAN (cGAN) structure is utilized to synthesize high-quality tumour images that imitate real data distributions more closely, improving data diversity and reducing overfitting. A mixture of Binary Cross-Entropy loss, L1 reconstruction loss, and adversarial loss are employed to train the GAN-OSMOSIS framework, balancing segmentation performance with fidelity of the output image [5]. By employing five principal performance metrics Dice Coefficient (0.92), Intersection over Union (IoU = 0.88), Sensitivity (0.91), Structural Similarity Index (SSIM = 0.94), and Fréchet Inception Distance (FID = 12.7) the system was thoroughly and extensively validated on five publicly available medical datasets. The new model also performed better than classical methods and showed higher generalization through 5-fold cross validation. The results confirm GAN-OSMOSIS as a solution for rare cancer segmentation, and present a scalable avenue for synthetic data augmentation in medical imaging research.