Object detection is central to histopathological image analysis, yet standard metrics (e.g. AP, F1-score) reveal little about where and why models fail, especially under domain shifts across scanners, tissues, or laboratories. In this study [1], we propose a quantitative, non-invasive framework to assess and class separability directly in the latent spaces of dense, fully convolutional object detectors with local correspondences. By aggregating object-level activations into size-invariant descriptors and computing separability per layer, our method reveals how discrimination emerges, propagates, or collapses within backbone, feature pyramid network (FPN), and detection heads. We adapt two metrics: an adapted generalized discrimination value (aGDV), contrasting inter- vs. intra-class distances, and a hellinger distance-based discrimination value (HDV), quantifying distribution overlap via the Bhattacharyya coefficient with variance-based channel selection and adaptive binning. On three real world datasets covering mitotic figure detection and multi-class cell detection, separability is modest in early backbone layers, increases through deeper backbone and FPN, and peaks in detection heads aligned with the relevant object scale. Layerwise HDV further differentiates domains with good generalization from those with severe domain shift and reflects class-specific confusion in multi-class settings. The framework provides actionable guidance for architecture and training decisions, indicating where discrimination is gained or lost, the impact of multi-domain training, stain augmentation, or self-supervised learning.