An Improved Nested U-Net Network for Fluorescence In Situ Hybridization Cell Image Segmentation

Zini Jian, Tianxiang Song, Zhihui Zhang, Zhao Ai, Heng Zhao, Man Tang, Kan LiuSensors 2024 · DOI

TL;DR

This paper introduces SEAM-Unet++, a novel deep learning model that enhances cell segmentation in challenging Fluorescence In Situ Hybridization (FISH) images. By integrating a specialized attention mechanism (SEAM) into the U-Net++ architecture, the model significantly improves the accuracy of identifying individual cell boundaries, especially when cells are clustered or overlapping.

What Problem Does This Paper Address?

Fluorescence In Situ Hybridization (FISH) is a crucial technique for detecting and localizing nucleic acid sequences, playing a vital role in medical diagnostics and biological research. However, analyzing FISH images presents significant challenges: the sheer number of cells, their disorganized arrangement, and the often-low resolution and noisy nature of the images make manual segmentation a tedious, time-consuming, and error-prone process. Traditional image processing methods struggle with the fine details and overlapping contours common in these images. Deep learning offers a promising solution, but achieving high accuracy in medical image segmentation, particularly for complex scenarios like adherent cells in FISH images, remains an open problem.

Methods

The core of this work is the SEAM-Unet++ model, which builds upon the U-Net++ architecture. U-Net++ is an extension of the U-Net model, designed to improve segmentation by introducing deeper nested skip connections and richer feature fusion mechanisms. This allows it to capture features at various scales more effectively.

The key innovation is the integration of the SEAM (Self-Attention Module) into the U-Net++ framework. SEAM is specifically designed to address issues like inter-class occlusion, local aliasing, and missing features, which are prevalent in FISH images with clustered cells. It works by enhancing the model's focus on important features and regions, particularly around occluded or weakly defined cell contours. The SEAM module is applied to the output of the U-Net++'s deepest layer (X0,4), allowing it to leverage the comprehensive features learned by the entire network. The SEAM module itself employs depth-separable convolutions and a simplified attention mechanism to efficiently learn relationships between features and adaptively weight them.

Additionally, the paper utilizes depth-separable convolutions within the SEAM module to reduce parameter count and computational complexity while learning spatial and channel correlations.

Datasets

The study utilized 246 FISH cell images acquired using an inverted fluorescence microscope at 60x magnification. These images were manually labeled using Labelme software, resulting in 199 labeled FISH cell images of size 1920 × 1080.

To address the limited dataset size, extensive data augmentation techniques were applied, including:

  • Cropping
  • Rotation
  • Mirroring
  • Raising grayscale value
  • Adding Gaussian noise

After augmentation, the dataset was processed into 2307 cell images, each with a size of 256 × 256. 80% of this augmented dataset was used for training, and the remaining 20% for testing.

Key Insights and Ideas

  • Attention for Adherent Cells: The SEAM attention mechanism is particularly effective at disentangling cells that are stuck together, a common and difficult problem in FISH image segmentation. It helps the model "see" individual contours even when they are physically touching.
  • Strategic Attention Placement: Instead of applying attention everywhere, the authors strategically place it at the deepest layer of U-Net++ to leverage rich, multi-scale features without risking overfitting.
  • Balancing Complexity and Performance: The use of depth-separable convolutions in SEAM helps maintain a good balance between model complexity and segmentation performance, crucial for efficient deep learning in medical imaging.
  • Data Augmentation is Crucial: The significant improvement in the dataset size through various augmentation techniques highlights its importance for training robust deep learning models on limited medical imaging data.

Key Results

The SEAM-Unet++ model demonstrated superior performance across several key metrics compared to other established segmentation models (Unet++, Unet, PSPNet, DeepLab, and Kun Lan's model).

  • Accuracy: Achieved the highest accuracy at 0.9153.
  • Loss: Recorded the lowest loss value at 0.1037.
  • IoU (Intersection over Union): Reached 0.9101, indicating excellent overlap between predicted and ground truth segmentations.
  • Recall: Achieved 0.8977, demonstrating a strong ability to identify all relevant cells without missing them.
  • Precision: Reached 0.8800, showing high accuracy in its positive predictions.
  • Dice Coefficient: Achieved an impressive 0.9432, signifying high similarity between the predicted and ground truth segmentation masks.

Qualitatively, the SEAM-Unet++ model produced significantly cleaner segmentation masks, especially for aggregated cells, with fewer instances of contour sticking compared to other methods.

Limitations and Future Work

  • Limitations: While SEAM-Unet++ shows significant improvement, the authors acknowledge that its performance in dealing with complex FISH images still needs to be improved. The paper also notes that the model was trained on a specific patch size (7) for the SEAM module, which might require further optimization for different datasets.
  • Future Work: The authors plan to expand the applicability of their model beyond FISH images to other medical imaging modalities, such as cardiac ultrasound and CT scan images. This involves exploring and enhancing the effectiveness of deep learning models across a wider spectrum of medical image types.

Summary updated

Abstract

Fluorescence in situ hybridization (FISH) is a powerful cytogenetic method used to precisely detect and localize nucleic acid sequences. This technique is proving to be an invaluable tool in medical diagnostics and has made significant contributions to biology and the life sciences. However, the number of cells is large and the nucleic acid sequences are disorganized in the FISH images taken using the microscope. Processing and analyzing images is a time-consuming and laborious task for researchers, as it can easily tire the human eyes and lead to errors in judgment. In recent years, deep learning has made significant progress in the field of medical imaging, especially the successful application of introducing the attention mechanism. The attention mechanism, as a key component of deep learning, improves the understanding and interpretation of medical images by giving different weights to different regions of the image, enabling the model to focus more on important features. To address the challenges in FISH image analysis, we combined medical imaging with deep learning to develop the SEAM-Unet++ automated cell contour segmentation algorithm with integrated attention mechanism. The significant advantage of this algorithm is that it improves the accuracy of cell contours in FISH images. Experiments have demonstrated that by introducing the attention mechanism, our method is able to segment cells that are adherent to each other more efficiently.
BibTeX
@inproceedings{jian2024improved,
  author = {Zini Jian and Tianxiang Song and Zhihui Zhang and Zhao Ai and Heng Zhao and Man Tang and Kan Liu},
  title = {{An Improved Nested U-Net Network for Fluorescence In Situ Hybridization Cell Image Segmentation}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24030928},
}
Figure