Segmentação geométrica de núcleos em Imagens microscópicas do desenvolvimento embrionário da Drosophila melanogaster

Abstract

Drosophila melanogaster is a widely used model organism in developmental biology, and microscopy image analysis is an essential tool for quantifying cellular patterns. In this context, this work presents a computer vision-based approach for nuclear instance segmentation in microscopy images of embryos, focusing on high-density scenarios with touching structures. The proposed method uses a geometric pipeline that models contours in polar representation 𝑟(𝜃) and detects concavities from radial variations with respect to a reference derived from the convex hull. Based on these concavities, directional alignment rules, geometric consistency criteria, and fallback mechanisms are used to determine separation cuts between possible nuclear instances. To support the application and validation of the approach, an open-source Python software with a graphical user interface was developed, integrating visual parameter tuning, batch processing, inspection of intermediate steps, and automatic metric extraction. The quantitative evaluation was conducted on a subset of images containing wild-type and mutant samples, using manually annotated reference masks (ground truth) spatially normalized through B-Spline interpolation (BOOR, 2001). The results indicated promising performance in nuclear mass delineation, reaching a Dice coefficient of up to 85.6% in the baseline model. At the instance level, the method showed variable performance across samples, with F1-Score values of up to 77.0%, indicating that the approach is promising for nucleus individualization in part of the evaluated scenarios, although it still presents limitations in cases of higher morphological complexity. Parameter calibration also showed potential for reducing geometric segmentation errors.