Nuclei Segmentation in Hepatocytes Using YOLO and SAM

• Published in: International Conference on Systems, Signals, and Image Processing (Online) pp. 1 - 6
• Main Authors: Da Silva, Sergio A., Berton, Leandro E. F., Carvalho, Mateus F. T., Bernardo, Carla C. O., Perles, Juliana V. C. M., Zanoni, Jacqueline N., Nanni, Loris, Sevilha, Andre L. R. G., Felipe, Gustavo Z., Flores, Franklin C., Costa, Yandre M. G.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09.07.2024
• Subjects: Analytical models; Annotations; Image segmentation; Manuals; Object detection; Quantification; Segment Anything Model; Segmentation; Task analysis; YOLO
• ISSN: 2157-8702
• DOI: 10.1109/IWSSIP62407.2024.10634020

This paper presents a hybrid approach that inte- grates the YOLO architecture for the detection of hepatocyte nuclei and the SAM for the segmentation of these cellular structures. This hybrid approach aims to facilitate and improve the task of quantification and morphometry analysis of cells in histopathological images, thereby providing researchers with a tool for automating their analysis without significant statistical difference from manual annotation. Three different versions of the YOLO model were tested: YOLOv3, YOLOv8, and YOLOv9. They achieved an AP of 70.28%, 70.98%, and 71.5% respectively. In terms of precision and recall, YOLOv3 achieved 90.7% and 92.4%, YOLOv8 achieved 90.4% and 92.94%, and YOLOv9 achieved 92.16% and 93.66%. According to the paired t-test, none of the YOLO models showed a significant statistical difference compared to human annotation. For morphometry analysis, we evaluated the default SAM and the fine-tuned models MED-SAM and Micro-SAM. The best results were obtained by the default SAM model, with a Dice coefficient of 93.65% and an IOU of 88.2%.