FCFDiff-Net: full-conditional feature diffusion embedded network for 3D brain tumor segmentation.

Background: Brain tumor segmentation (BraTS) plays a critical role in medical imaging for early diagnosis and treatment planning. Recently, diffusion models have provided new insights into image segmentation, achieving significant success due to their ability to model nonlinearities. However, existing methods still face challenges, such as false negatives and false positives, caused by image blurring and noise interference, which remain major obstacles. This study aimed to develop a novel neural network approach to address these challenges in three-dimensional (3D) BraTS.

Methods: We propose a novel full-conditional feature diffusion embedded network (FCFDiff-Net) for 3D BraTS. This model enhances segmentation accuracy and robustness, particularly in noisy or ambiguous regions. This model introduces the full-conditional feature embedding (FCFE) module and employs a more comprehensive conditional embedding approach, fully integrating feature information from the original image into the diffusion model. It establishes an effective connection between the decoder side of the denoising network and the encoder side of the diffusion model, thereby improving the model's ability to capture the tumor target region and its boundaries. To further optimize performance and minimize discrepancies between conditional features and the denoising module, we introduce the multi-head attention residual fusion (MHARF) module. The MHARF module integrates features from the FCFE with noisy features generated during the denoising process. Using multi-head attention aligns semantic and noise information refining the segmentation results. This fusion enhances segmentation accuracy and stability by reducing noise impact and ensuring greater consistency across tumor regions.

Results: The BraTS 2020 and BraTS 2021 datasets served as the primary training and evaluation datasets. The proposed architecture was assessed using metrics such as Dice similarity coefficient (DSC), Hausdorff distance at the 95th percentile (HD95), specificity, and false positive rate (FPR). For the BraTS 2020 dataset, the DSC scores for the whole tumor (WT), tumor core (TC), and enhancing tumor (ET) were 0.916, 0.860, and 0.786, respectively. The HD95 values were 1.917, 2.571, and 2.581 mm, whereas specificity values were 0.998, 0.999, and 0.999, and FPR values were 0.002, 0.001, and 0.001, respectively. On the BraTS 2021 dataset, the DSC scores for the same regions were 0.926, 0.903, and 0.869, with HD95 values of 2.156, 1.834, and 1.583 mm, respectively. Specificity and FPR values were 0.999 across the board, and FPR values were consistently low at 0.001. These results demonstrate the model's excellent performance across the three regions.

Conclusions: The proposed FCFDiff-Net provides an efficient and robust solution for 3D BraTS, outperforming existing models in terms of accuracy and robustness. Future work will focus on exploring the model's generalization capabilities and conducting lightweight experiments to further enhance its applicability.