Neurosymbolic approach improves robustness of deep learning models for semantic segmentation of MII human oocyte images

Morphology-based assessment of metaphase II (MII) oocytes remains an important component of intracytoplasmic sperm injection (ICSI) workflows, but it is affected by inter-operator variability and limited reproducibility. Deep learning approaches have shown promise for automated segmentation of oocyte structures, although purely data-driven models may generate anatomically implausible masks. Domain knowledge can instead be integrated into neural networks as explicit structural and anatomical constraints guiding training. We retrospectively analyzed 1,243 microscopic images acquired at 200× magnification from 218 patients undergoing ICSI at a single center. We propose a neurosymbolic architecture in which a logic score, derived from a set of knowledge-driven anatomical constraints, is computed on the regions of interest predicted by a transformer-based model (SegFormer). This score was used during training, alongside the segmentation loss, to guide parameter updates, and at inference as a continuous measure of morphological reliability. The neurosymbolic model was compared with an architecturally identical baseline, both quantitatively and qualitatively. To verify that the improvement is attributable to the content of the logical rules rather than to the presence of an auxiliary term, we trained two controls with identical loss form and weight but no anatomical meaning: a non-structural auxiliary loss and a scrambled rule set. The framework was further compared with U-Net and DeepLabV3+ trained on the same dataset and split, and the logic loss was additionally applied to a convolutional backbone. The neurosymbolic model achieved a macro-average intersection over union (IoU) of 0.84 and a Dice score of 0.91, outperforming the baseline (IoU 0.80; Dice 0.88), U-Net (IoU 0.76) and DeepLabV3+ (IoU 0.81). A non-structural auxiliary loss of equal weight did not reproduce this gain (IoU 0.81; Dice 0.88), and scrambling the semantics of the rules degraded performance (IoU 0.61), indicating that the improvement is driven by the content of the constraints rather than by the auxiliary term itself. By embedding domain knowledge as logical constraints, the proposed neurosymbolic framework yields more accurate and morphologically coherent segmentations of MII oocyte images than a purely neural SegFormer baseline. This approach represents a promising image-analysis component for future decision-support systems in ICSI workflows.