Image Processing of Porous Silicon Microarray in Refractive Index Change Detection

Zhiqing Guo, Zhenhong Jia, Jie Yang, Nikola K. Kasabov, Chuanxi LiSensors 2017 · OpenReview

TL;DR

This paper introduces a novel image processing pipeline for analyzing porous silicon (PSi) microarrays, specifically designed to overcome challenges posed by reflected light images. The method offers improved accuracy and speed in detecting refractive index changes, crucial for label-free biosensing applications.

What Problem Does This Paper Address?

Current commercial microarray detection methods, often relying on fluorescence, suffer from high costs, complex protocols, and potential interference with biological molecule activity. Label-free detection methods are gaining traction, with porous silicon (PSi) microarrays showing promise due to their sensitivity to refractive index changes. However, analyzing the reflected light images from PSi microarrays presents unique challenges:

  • Complex Background and Noise: Reflected light images are prone to speckle and impulse noise due to the interaction of laser light with the rough PSi surface.
  • Irregular Cell Structures: Electrochemical etching during PSi fabrication can lead to irregular cell edges, complicating segmentation.
  • Varying Reflectivity: Differences in photonic structures between the cell edge and center result in non-uniform reflectivity, further complicating image analysis.
  • Illumination Sensitivity: Standard RGB color spaces are susceptible to illumination variations, making direct binarization difficult.

Existing image processing algorithms, often developed for fluorescent images, are not well-suited for these specific challenges. This paper aims to develop a robust and efficient image processing pipeline for PSi microarray reflected light images to enable accurate and fast refractive index change detection.

Methods

The proposed method comprises three main stages: pretreatment, tilt correction, and spot segmentation.

  • Pretreatment:
    • HSV Color Space Conversion: The RGB image is converted to the HSV (Hue, Saturation, Value) color space. This is crucial because the HSV space separates color information (Hue, Saturation) from luminance (Value), making it less susceptible to illumination changes.
  • Component Processing: The Saturation and Value components are individually processed.
  • Saturation Component: Used to extract cell contours and suppress bright objects connected to the boundary.
  • Value Component: Used to represent cell size and is processed to remove irrelevant noise.
  • Binarization: The "maximum between class variance" method (Otsu's method) is employed for binarization.
  • Morphological Operations: Operations like "open operation" and hole filling are used to refine the binarized images and remove noise.
  • Tilt Correction:
    • Minimum Bounding Rectangle (MBR) Approach: A novel tilt correction algorithm based on MBR is proposed. This method involves:
  • Scanning the binarized image horizontally and vertically to determine an initial external rectangle.
  • Using the centroid of this rectangle as the center of rotation.
  • Iteratively rotating the principal axis within an acute angle area to find the MBR.
  • Calculating the tilt angle based on the transformation from the initial rectangle to the MBR.
  • This approach is designed to be computationally efficient compared to traditional methods like Radon transform.
  • Spot Segmentation:
    • Ellipse Fitting: After tilt correction and binarization, ellipse fitting is applied to the cell contours to obtain smooth boundaries.
  • Centroid Determination: The centroid coordinates of each cell are determined.
  • Regularized Dot Placement: To ensure consistent analysis, the algorithm enforces equal spacing and radius for the segmented dots. This involves:
  • Discretizing the horizontal axis to ensure equal spacing between dots in a row.
  • Averaging ordinates to determine the vertical position of each row.
  • Calculating the radius based on the horizontal and vertical distances of the cell, with a slight reduction (multiplying by 0.9) to exclude bright spots on the cell edges.

Datasets

The experiments were conducted using eighteen pictures of PSi microarrays. These pictures were divided into two categories based on their size and the angle of the incident laser:

  1. 6x6 arrays: Average size of 283x287 pixels, with reflection angles from 0° to 8°.
  2. 12x12 arrays: Average size of 567x576 pixels, with reflection angles from 0° to 8°.

The paper does not introduce new benchmark datasets but uses these experimental images to evaluate the proposed algorithms.

Key Insights and Ideas

  • HSV for Robustness: Leveraging the HSV color space, particularly the Saturation and Value components, is key to handling the illumination variations and complex backgrounds inherent in reflected light PSi microarray images.
  • MBR for Efficient Tilt Correction: The MBR-based tilt correction offers a significant speed advantage over traditional methods like Radon transform, making it suitable for real-time or large-scale analysis.
  • Ellipse Fitting for Smooth Boundaries: Using ellipse fitting after binarization provides a robust way to define cell boundaries, even when the original etched structures are irregular.
  • Regularized Segmentation for Consistent Analysis: Enforcing uniform spacing and radius for segmented dots is crucial for quantitative analysis of refractive index changes across different cells and experiments.

Key Results

  • Pretreatment: The proposed pretreatment method successfully eliminated background and light intensity interference, effectively segmenting the cells.
  • Tilt Correction: The MBR correction algorithm demonstrated good applicability and a significant advantage in computing time compared to the Radon transform.
  • For 6x6 arrays, the MBR correction had an average error of 0.16° and an average time of 1.13 s, compared to 0.17° and 8.61 s for Radon transform.
  • For 12x12 arrays, the MBR correction had an average error of 0.06° and an average time of 1.70 s, compared to 0.10° and 34.38 s for Radon transform.
  • The algorithm showed smaller error fluctuations than Radon transform.
  • Spot Segmentation: The segmentation algorithm effectively eliminated bright spots on cell edges and segmented dots into the relevant information region.
  • The average segmentation time from RGB image to completion for 6x6 arrays was 1.51 s, and for 12x12 arrays was 3.88 s.
  • The total average time for dots extraction (including tilt correction) for 6x6 arrays was 1.79 s, and for 12x12 arrays was 4.71 s.

Limitations and Future Work

  • Limitations:
    • The paper focuses on reflected light images from PSi microarrays. While it mentions the algorithm also has a good effect on fluorescent images, its primary optimization is for reflected light.
  • The effectiveness of the "maximum between class variance" method for binarization might still be limited in cases of extremely low signal-to-noise ratios, although the HSV preprocessing aims to mitigate this.
  • Future Work:
    • The authors suggest that the proposed method can be utilized for fast, accurate, and automatic dots extraction of PSi microarray reflected light images, implying potential for further refinement and application in real-world biosensing scenarios.
  • The paper's success in improving accuracy and speed suggests its suitability for developing more advanced, high-throughput PSi-based biosensing platforms.

Summary updated

BibTeX
@inproceedings{guo2017image,
  author = {Zhiqing Guo and Zhenhong Jia and Jie Yang and Nikola K. Kasabov and Chuanxi Li},
  title = {{Image Processing of Porous Silicon Microarray in Refractive Index Change Detection}},
  booktitle = {Sensors},
  year = {2017},
  url = {https://openreview.net/forum?id=UqIbiyAzad},
}
Figure