Preserving Fine-Grained Droplet Structures by Spatial Gaussian Blurring Against Pooling Deformation in U-Net-Based Semantic Segmentation

Pooling operations in convolutional neural networks often lead to the loss of fine-grained object features, particularly when those objects occupy only a few pixels (e.g., 5 × 5 or smaller). This problem is critical in droplet analysis, where pooling-induced feature loss causes deformation of fine-grained droplet boundaries within the encoded representation. When fine-grained droplet features are structurally deformed and vanished in the encoder, the resulting ineffective representation impairs learning and leads to inaccurate droplet segmentation, which severely impairs downstream decision-making. We propose a data augmentation strategy using Gaussian blurring to mitigate this problem. By applying spatial blurring prior to encoding, the blur acts as a contour regularization, strengthening the network learning capability to retain fine-grained droplet structure through subsequent pooling layers and improving robustness against pooling deformation. To further strengthen spatial feature interaction, we integrate a graph convolution module at the bottleneck of a U-Net architecture. The graph is constructed from the bottleneck feature map, with edges weighted by raw intensity values of the corresponding feature elements. This non-parametric, intensity-based graph formulation captures structural relationships among features and complements spatial feature preservation. Our combined method achieves a 16.5% relative improvement in Recall (R), indicating that more fine-grained droplets are correctly segmented compared to the baseline U-Net. Ablation studies confirm that Gaussian blur augmentation is the dominant contributor to these gains, while the GraphSAGE provides complementary refinement. The code is available at: https://github.com/lynerlwl/blur-unet-segmentation