UCTLFANet: a low-rank fine-tuning model for microalgae image segmentation

Area of Science:

• Marine Biology
• Computational Biology
• Image Analysis

Background:

• Accurate microalgae image segmentation is crucial for morphological analysis, cell counting, and physiological assessment.
• Existing models often require significant computational resources for fine-tuning, limiting their efficiency.

Purpose of the Study:

• To introduce UCTLFANet, a novel low-rank fine-tuning model for microalgae image segmentation.
• To optimize parameter efficiency and computational resource conservation during model fine-tuning.
• To enhance the accuracy and applicability of microalgae image analysis.

Main Methods:

• Proposed a low-rank fine-tuning method using a large language model (LLM) integrated with the LoRA-FA technique.
• Applied the fine-tuning module to the UCTLFANet architecture, specifically testing integration into fully connected and convolutional layers.
• Compared the performance of UCTLFANet against established models like UNet, UNet++, Attention-UNet, and UCTransNet.

Main Results:

• Integrating the LoRA-FA module into the fully connected layer of UCTLFANet yielded the best segmentation performance.
• UCTLFANet demonstrated superior accuracy and efficiency compared to UNet, UNet++, Attention-UNet, and UCTransNet.
• The fine-tuned UCTLFANet showed strong potential for generalization across different datasets and applications in marine microbial imaging.

Conclusions:

• The proposed low-rank fine-tuning method significantly enhances microalgae image segmentation efficiency and accuracy.
• UCTLFANet offers a computationally efficient and effective solution for microalgae analysis.
• The model's adaptability suggests broad utility in marine microbial imaging and related fields.