Author: Kota Hirose, Daisuke Kawahara, Jokichi Kawazoe, Yuji Murakami 👨🔬
Affiliation: Department of Radiation Oncology, Graduate School of Biomedical and Health Sciences, Hiroshima University, Graduate School of Biomedical and Health Sciences, Hiroshima University 🌍
Purpose: Synthesizing medical images can address the lack of or unscanned medical images, reducing scanner time and costs. However, paired image scarcity remains a challenge for image synthesis. We propose a novel unpaired image generation model incorporating contrastive learning and a hybrid CNN-Transformer structure to enhance medical image generation. For segmentation evaluation, we leverage the generalizability of the Segment Anything Model 2 (SAM2), trained on large-scale datasets. The proposed approach is validated using real and synthesized MR images of cerebral infarction, assessed through image quality metrics and lesion segmentation performance.
Methods: We used a public dataset comprising DWI and ADC MR images, and masks of cerebral infarction regions from 250 patients with cerebral infarction. The proposed model integrates CNN and Transformer into the Contrastive unpaired translation framework, enabling simultaneous capture of local features and global contexts within images. Using this model, ADC images were synthesized from DWI images. The synthesized images were evaluated using the Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR). For lesion detectability, SAM2 was incorporated into a U-Net structure to create a SAM2-U-Net model for automated segmentation. Lesion detection performance was assessed using the Dice similarity coefficient (DSC) and Intersection over Union (IoU).
Results: The SSIM and PSNR of the synthesized ADC images were 0.96 and 29.6, respectively. Lesion detectability using real ADC images achieved a DSC of 0.82 and IoU of 0.69, while synthesized ADC images showed a DSC of 0.80 and IoU of 0.67, with a difference of less than 0.02. Previous models with conventional U-Net architecture achieved a DSC of 0.60, showing significant improvement with our hybrid model.
Conclusion: The proposed method successfully synthesized MR images with comparable quality metrics and lesion detectability to real MR images. Furthermore, the integration of SAM2 enabled highly accurate automated lesion detection.