Lifelong learning enabled GAN framework for brain MRI FLAIR synthesis and tumor segmentation

Accurate synthesis of missing MRI modalities plays a crucial role in multi-modal neuro-imaging pipelines. In real-world setups, each hospital or scanner does not always provide all MRI modalities for diagnosis. As FLAIR MRI sequence shows tumor boundaries especially edema, many sites lack FLAIR MRI sequence or have low-quality FLAIR. In this work, we propose a 2.5D (using 7 slices) GAN-based framework for synthesizing missing modality i.e., FLAIR images from T1-weighted and T2-weighted inputs, followed by tumor segmentation on multiple sequential tasks, each representing a new data distribution. The generator employs a U-Net architecture enhanced with residual blocks and convolutional block attention module (CBAM), while a PatchGAN discriminator and an U-Net-based segmentor to ensure both visual fidelity and downstream utility. The model is trained with a combination of perceptual, structural losses, along with segmentation consistency constraints on a lifelong learning strategy using EWC method on 4 sequential tasks from BraTS2020 dataset. Quantitatively, the proposed model achieves an average SSIM score of 0.902, PSNR score of 28.375 dB, MAE value of 0.023 and LPIPS value of 0.119 over sequential tasks. The model has also achieved an average Dice and HD95 score of 0.79 and 7.68 mm respectively. The model shows very good performance on backward as well as forward transfer of knowledge, which confirm the effectiveness of synthesis-segmentation pipeline and the potential of lifelong learning in medical image synthesis and segmentation with good accuracy and minimal catastrophic forgetting.