A New Deep Learning Process For Postoperative Corneal Topography Prediction Based On Raw Multimodal Data From The Pentacam Hr System

This work addresses the limitations of existing GAN models, particularly their inability to effectively capture global feature relationships due to the locality of convolutional neural network (CNN)-based approaches, while mitigating challenges such as image noise and geometric information loss during training.

Shenzhen Eye Hospital, Shenzhen Eye Center, Southern Medical University

The MCSTransWnet framework integrates a generator and a discriminator. By fusing transformer-based global feature extraction with U-net’s local feature preservation, the model simultaneously captures contextual relationships and fine-grained image details. This marks the first integration of transformers, U-nets, and multi-conditional modules in a GAN framework for medical topography prediction.

Experimental evaluations demonstrate the model’s effectiveness, achieving a structural similarity index (SSIM) of 0.77, peak signal-to-noise ratio (PSNR) of 16.03 dB, and Fréchet inception distance (FID) of 9.11. These metrics confirm the model’s capability to generate clinically relevant postoperative corneal topographies with high structural fidelity and low distortion.

The MCSTransWnet model successfully bridges the gap between global feature modeling and local detail preservation in medical image prediction tasks. Its novel fusion of transformers and U-nets within a conditional GAN framework outperforms conventional CNN-based approaches, offering a robust solution for surgical outcome simulation. This work establishes a new paradigm for multimodal medical data integration and holds potential for enhancing preoperative planning in refractive surgeries.