Existing automated deep learning frameworks for cellular analysis, such as the standard U-Net and Fully Convolutional Networks (FCN), face significant performance degradation when processing clinical samples with high morphological variance, dense clumping, and low-contrast borders. While advanced iterations like Transformer-based U-Nets and pure Kolmogorov-Arnold Networks (KANs) attempt to address these limitations, they introduce substantial technical gaps. Transformers suffer from prohibitive computational overhead and demand massive annotated datasets, whereas pure spline-based models are highly sensitive to noise and introduce severe memory strain. Consequently, existing automated models struggle to maintain fine-grained boundary definitions without suffering from edge-blurring pitfalls or structural inefficiency in aggregated cell environments.To address these limitations, this study evaluates an improved and more effective Kolmogorov-Arnold Attention U-Net (KC-UNet) architecture. This framework integrates learnable B-spline activation functions directly into a localized two-dimensional latent bottleneck to resolve overlapping junctions, while utilizing Convolutional Block Attention Modules along the skip connections to suppress background noise and enhance multi-scale feature routing. The performance of this system was evaluated against standard baseline models, including U-Net, FCN, and DeepLabV3, using a rigorous nested 5-fold cross-validation strategy across both a static original dataset and an augmented dataset injected with simulated clinical artifacts.The empirical results demonstrate a notable improvement in both performance resilience and clinical utility. Under dataset augmentation, standard models suffered severe performance drop-offs, with U-Net and DeepLabV3 yielding negative mean Intersection over Union (mIoU) delta changes. In contrast, the KC-UNet was the only architecture to maintain structural stability under clinical noise, demonstrating a positive resilience change of +0.0133 mIoU. Downstream instance counting accuracy further validated its clinical value, with the architecture achieving the lowest Mean Absolute Error (MAE) of 13.15, significantly outperforming the standard U-Net (MAE of 23.50) and DeepLabV3 (MAE of 30.40). Validation analysis additionally confirmed that this framework provides the narrowest error variance and a 50% reduction in systematic counting bias compared to baseline convolutional models.This research provides a highly robust, parameter-efficient framework for automated hematological diagnostics, demonstrating that combining the local feature-extraction capabilities of convolutions with the non-linear flexibility of learnable spline edges resolves dense platelet clumping effectively. By mitigating operator bias and batch-dependent manual threshold adjustments, this strategy advances automated high-throughput clinical screening pipelines. Future work will focus on structural model pruning for embedded clinical deployment, accelerating edge computations, and exploring self-supervised learning paradigms to eliminate the pipeline's reliance on pseudo-ground truth masks.