Difference-map-enhanced dice loss for fetal brain segmentation

Congenital anomalies are one of the primary causes of infant mortality, making early detection during pregnancy crucial. Accurate fetal brain segmentation in MRI is essential for evaluating these abnormalities, but manual methods are time-consuming and subject to variability. Deep Learning approaches proved to generate accurate segmentations, but the choice of loss function is crucial for optimizing model training and directly influences the model’s ability to learn and make precise predictions. This study introduces a novel loss function denominated Difference Map-Enhanced Dice Loss (DD-Map loss) designed to enhance segmentation accuracy for fetal brain MRI. The methodology consisted in generating a mismatch map between the predictions and the ground-truth and subsequently applying a Gaussian smoothing to create a smooth map of differences. The final loss function combined a sigmoid-weighted L1 loss for minimizing this map with the Dice loss to improve overall segmentation performance and refine regions close to the contours. The U-Net model was used to test this new loss function against established ones, including Dice, Tversky, Focal, Cross-Entropy (CE), Hausdorff Distance (HD), Generalized Dice (GD) and other compound losses. Training and evaluations were conducted on the FeTA (Fetal Tissue Annotation) dataset, which consists of 80 3D MRI images. The results showed that the proposed loss improved the segmentation task, with the model achieving a Dice Similarity Coefficient (DSC) of 84.83% and an Average Surface Distance (AverSSD) of 0.99mm, outperforming other methods. The new loss demonstrated strong performance, particularly on larger, more clearly defined brain structures. Although it did not show significant improvements for smaller structures, the observed differences were not statistically significant compared to the other losses. Thus, the loss function exhibited superior accuracy and robustness, showing potential to be used to improve fetal brain MRI segmentation.