Enhancing DL-based Cell Segmentation of Microalgae with Classical Image Processing Priors

We present a hybrid approach for the automatic de- tection and segmentation of Nannochloropsis Oceanica, Wild Type (NocWT) microalgae cells, combining classical image processing techniques with deep learning. Initially, we apply traditional computer vision methods to detect and count cells efficiently, but these struggle with challenges such as morphological variability and overlapping structures. To overcome these limitations, we incorporate the Segment Anything Model (SAM), a state-of-the- art segmentation framework leveraging a transformer archi- tecture pre-trained on large-scale datasets. Instead of relying solely on SAM’s general capabilities, we guide its segmentation using pre-segmented regions derived from classical methods, improving accuracy in delineating complex cell boundaries. The proposed method is evaluated on a manually annotated dataset of bright-field microscopic images, ensuring reliable performance assessment despite the dataset’s limited size. By integrating the interpretability of traditional approaches with the adaptability of deep learning, our method achieves robust and precise microalgae segmentation, demonstrating the advantages of a complementary strategy over standalone state-of-the-art techniques.