A Lightweight Modified Adaptive UNet for Nucleus Segmentation.

Area of Science:

• * Computational biology
• * Biomedical image analysis
• * Deep learning for microscopy

Background:

• * Cell nucleus segmentation is vital for quantitative analysis in biological and biomedical research.
• * Traditional methods and existing U-Net models struggle with small foreground objects and imbalanced datasets.
• * Large model sizes in current methods lead to overfitting issues.

Purpose of the Study:

• * To develop a novel deep learning architecture, mA-UNet, for improved cell nucleus segmentation.
• * To address challenges in predicting small foreground elements and dataset imbalance.
• * To evaluate the performance of mA-UNet against state-of-the-art models.

Main Methods:

• * Introduction of the mA-UNet architecture, optimized for small foreground element prediction.
• * Implementation of a data preprocessing strategy adapted from road segmentation to handle imbalanced datasets.
• * Quantitative and qualitative evaluation on the 2018 Data Science Bowl dataset.

Main Results:

• * mA-UNet achieved a Mean Intersection over Union (MIoU) score of 95.50%, outperforming UNet++.
• * The proposed methodology demonstrated superior performance compared to other state-of-the-art models.
• * mA-UNet was successfully implemented on Zynq UltraScale+ FPGA using VHDL, showing efficiency and scalability.

Conclusions:

• * mA-UNet offers a significant advancement in cell nucleus segmentation for microscopy images.
• * The model effectively handles small objects and imbalanced data, crucial for biological applications.
• * The FPGA implementation highlights the model's efficiency and potential for real-time applications.