Research and Analysis of Fine-tuning Techniques for Cell Image Segmentation Model Based on SAM2

Abstract

This study proposes a specialized framework for finetuning the Segment Anything Model 2 (SAM2) to address the challenging task of biomedical blood cell image and video segmentation. To overcome issues like poor contrast, blurred boundaries, and cell adhesion, the paper curated a hybrid dataset of 400 annotated images from the public LISC database and an in-house collection. A standardized preprocessing pipeline involving CLAHE and Z-score normalization was applied. The parameterefficient fine-tuning strategy selectively unfroze the final layers of the image encoder and the entire memory encoder, incorporating Low-Rank Adaptation (LoRA) and lightweight adapters. Training was guided by a composite loss function combining segmentation (Focal and Dice losses), temporal consistency, and morphological regularization terms. Experimental results show that fine-tuned model, ‘BloodCellSAM2’, achieves a mean Intersection over Union (mIoU) of 85.7% and a Dice coefficient of 91.3% on the test set, representing a significant improvement over the original SAM2 (mIoU: 65.2%) and a U-Net baseline (mIoU: 78.4%). This approach delivers high accuracy while training only 0.8% of the parameters, offering an efficient and robust solution for automated cell analysis in clinical and research settings.