Ciro Russo

Background and Objective: Accurate nuclei detection is a fundamental prerequisite for automated computational cytology pipelines supporting cancer screening and routine cytological assessment. Robust detection remains challenging because of high variability in staining appearance, cellular morphology, and imaging conditions across cytological domains. Most existing methods adopt object detection paradigms originally developed for natural images and are typically evaluated within a single cytological domain, leaving the behavior of pixel-wise detection formulations across heterogeneous settings largely uncharacterized. This study aims to systematically characterize a pixel-wise detection formulation based on gravity-point representations for nuclei localization across multiple, heterogeneous cytology domains.Methods: We evaluated a pixel-wise formulation that places candidate points on a regular grid over a backbone feature map, assigning each point a classification score and a regression offset toward the nearest nucleus center, with final detections obtained by non-maximum suppression. The formulation was assessed on three publicly available datasets covering cervical, oral, and nasal cytology, differing in staining protocol, cellular morphology, nuclear density, and acquisition conditions. We analyzed the influence of the feature extraction backbone and of the gravity-point sampling configuration, and benchmarked the formulation against representative anchor-based, anchor-free, and transformer-based detectors under a unified experimental protocol. Detection performance was quantified using precision, recall, and F1-score at a fixed operating point, complemented by threshold-independent curve-based metrics.Results: Sensitivity to backbone selection was limited across all domains, whereas the gravity-point sampling configuration introduced more pronounced variations whose magnitude depended on domain characteristics such as nuclear density and morphological heterogeneity. The formulation achieved F1-scores of 94.5%, 77.1%, and 87.3% on the cervical, oral, and nasal datasets, respectively, with recall values of 96.9%, 75.5%, and 81.6% and precision values of 92.2%, 78.7%, and 93.8%. Performance was comparable to or higher than the representative object detection paradigms, with the most pronounced advantages in recall and in the low false-positive regime.Conclusions: A pixel-wise, point-based formulation provides a consistent and competitive approach to nuclei localization across heterogeneous cytological domains, with the candidate sampling configuration emerging as the more relevant design factor. These findings support point-based detection as a robust building block for automated cytology pipelines and motivate future work on adaptive sampling and validation in real diagnostic workflows.