Background and Objective: Generating realistic medical images along with anatomically accurate segmentation masks is a promising approach to address the shortage of annotated data in medical imaging, particularly in optical coherence tomography (OCT) of mouse eyes, where manual retinal layer delineation is very labour-intensive due to tiny structures and demands on expert knowledge, resulting in scarce available datasets. While diffusion models have shown strong performance in medical image synthesis, joint image-mask generation has primarily relied on U-Net-based denoisers, leaving diffusion transformers largely unexplored in this context.Methods: We propose a conditional dual-output Diffusion Transformer (DualDiT) for the simultaneous synthesis of OCT B-scans and segmentation masks of the upper retinal cell layers (URCL) of ex vivo mouse retina. DualDiT encodes both modalities into a shared latent space via a pretrained variational auto-encoder, concatenates their latent representations, and performs conditional diffusion over the resulting joint tensor. We compared DualDiT against two adapted denoising diffusion baselines: a pixel-space conditional model (DDPM) and a latent-space conditional model (LDM). Generative quality was assessed using distribution similarity metrics (Fréchet Inception Distance, FID; spatial FID, sFID), practical utility was evaluated through synthetic data augmentation for downstream U-Net segmentation, and perceptual realism was assessed by a panel of three domain experts.Results: DualDiT achieved the best generative quality, obtaining a FID of 56.14 and sFID of 114.35, outperformingDDPM(FID 164.55, sFID 254.52) andLDM(FID 102.21, sFID 150.66). In the expert evaluation, panels incorrectly classified, on average, 46% of synthetic samples as real and 42% of real samples as synthetic. The additional use of DualDiT-generated images and masks increases Dice and IoU scores, as evaluated on a held-out test set of the segmentation model.Conclusions: DualDiT demonstrates that transformer-based diffusion models can effectively learn the joint distribution of OCT images and segmentation masks, surpassing conventional DDPM- and LDM-based baselines in generative fidelity, downstream segmentation utility, and perceptual realism, highlighting its potential as a data augmentation strategy for annotation-scarce medical imaging applications.