Pathological Myopia (PM) progresses through distinct stages - Tessellated Fundus, Choroidal Atrophy, and Patchy Atrophy. The limited availability of annotated datasets poses challenges for developing machine learning models tailored to these stages. This study introduces a novel framework for synthetic image generation using CycleGAN, enhanced with a hybrid perceptual loss. By leveraging a CNN-based feature extractor, this approach refines biomarker representation and ensures their preservation across PM stages while improving image quality. The hybrid perceptual loss aligns generated images with high-level features from real images, enhancing biomarker accuracy and structural fidelity. This methodology not only augments dataset diversity but also facilitates clinical applications by producing synthetic images that faithfully represent PM stages and biomarkers, contributing to the advancement of ophthalmological diagnostics.