Department of Computer Science and Engineering, College of Engineering, Ewha Womans University, Seoul 03760, Republic of Korea.
Division of Biomedical Engineering, College of Health Sciences, Yonsei University, Wonju, Gangwon-do 26493, Republic of Korea.
Micron. 2019 Nov;126:102718. doi: 10.1016/j.micron.2019.102718. Epub 2019 Aug 16.
The morphology of tumor cells is highly related to their phenotype and activity. To verify the drug response of a brain tumor patient, fluorescence microscope images of drug-treated patient-derived cells in each well are analyzed. Due to the limitation of the field of view (FOV), a large number of small FOVs are acquired to compose one complete microscope well. Here, we propose an automated method for accurately stitching tile-scanned fluorescence microscope images, even with noise and a narrow overlapping region between adjacent fields. The proposed method is based on intensity-based normalized cross-correlation (NCC) and a triangular method-based threshold. The proposed method's quantitative accuracy and the sensitivity of the input was compared to other existing stitching tools, MIST and FijiIS, setting manually stitched images as the ground truth. The test images were 20 samples of 3 × 3 grid images in three versions of the fluorescence channel. The distance between the location of each field and number of cells was determined for different input field overlap ranges (1%, 3%, 5%, and 10%), while the actual value was about 1.15%. The proposed method had a distance error of 1.5 pixels at an input overlap of 1%, showing the lowest minimum error at all channels. Regarding the difference in cell numbers, although the number of overlapping cells was always small because of the narrow overlapping range, the proposed method was able to generate the resultant image with the smallest difference. In addition, to confirm the size limitation of the proposed algorithm, the accuracy of stitching images of grid structures 3 × 3, 5 × 5, 10 × 10-20 × 20 was tested, showing consistent results. In conclusion, quantitative evaluation of the performance of the method proved its improved accuracy compared to other current state-of-art techniques, and it showed robust performance even with noise and a narrow overlapping region between adjacent fields.
肿瘤细胞的形态与其表型和活性高度相关。为了验证脑肿瘤患者的药物反应，需要分析每个孔中药物处理的患者来源细胞的荧光显微镜图像。由于视场（FOV）的限制，需要采集大量小的 FOV 来组成一个完整的显微镜孔。在这里，我们提出了一种自动拼接 tile-scanned 荧光显微镜图像的方法，即使在存在噪声和相邻场之间的重叠区域较窄的情况下也能准确拼接。该方法基于基于强度的归一化互相关（NCC）和基于三角形的阈值。将手动拼接图像作为ground truth，将该方法的定量准确性和输入的灵敏度与其他现有的拼接工具 MIST 和 FijiIS 进行了比较。测试图像是三种荧光通道的 20 个 3×3 网格图像样本。对于不同的输入场重叠范围（1%、3%、5%和 10%），确定了每个场的位置和细胞数量之间的距离，而实际值约为 1.15%。在输入重叠为 1%的情况下，该方法的距离误差为 1.5 像素，在所有通道中显示出最小的最小误差。关于细胞数量的差异，尽管由于重叠范围较窄，重叠细胞的数量总是很小，但该方法能够生成具有最小差异的结果图像。此外，为了确认该算法的大小限制，还测试了 3×3、5×5、10×10-20×20 网格结构的图像拼接的准确性，结果一致。总之，对该方法性能的定量评估证明了其与其他当前最先进技术相比具有更高的准确性，并且即使在噪声和相邻场之间的重叠区域较窄的情况下也表现出稳健的性能。
