ALGORITHM FOR AUTOMATIC CREATION OF SEGMENTATION MASK FOR DETECTION OF BIOLOGICAL OBJECTS

Abstract

The article presents a method for automatically creating segmentation masks for biomedical images, which significantly reduces the laboriousness of manual annotation and increases the reproducibility of data preparation. The proposed approach combines adaptive thresholding with Gaussian matrix coefficients, morphological operations, and geometric filtering of contours by area and roundness coefficient. This combination allows for effective separation of cellular structures under conditions of uneven illumination, noise, and low contrast, which are typical problems of microscopic images. The method was tested on the BBBC030v1 dataset, which contains 60 images of Chinese hamster ovary cells. For each image, the automatically created mask was compared with the provided ground truth annotation using the Dice coefficient. The average value was 0.8954, the median was 0.9013, and the standard deviation was 0.0254, which indicates high accuracy and stability of the method. The narrow interquartile range (IQR = 0.0215) confirms the uniformity of the algorithm's performance on most samples, while single outliers (0.80–0.85) are associated with atypical or low-contrast images. The overall result demonstrates that the classical segmentation approach without the use of neural networks can achieve quality comparable to manual expert labeling. To verify the practical suitability of the generated masks, they were used to train the U-Net neural network for the segmentation task. Comparison with training on real masks showed almost identical results (0.9036 vs. 0.9037), which confirms the possibility of full or partial replacement of manual annotation by an automatic approach. The developed method can be applied to accelerate the preparation of large biomedical datasets and integration into decision support systems in cytology, histology and other fields of biomedicine.