Visualization in Biology and Medicine - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 240

or TIFF file format. Depending on the application and the microscope manufacturer,
the images have varying color depths and meta information, adding to the problem
of data format diversity. To address this issue, the BioFormats group has developed
a Java library for the reading and writing of a large variety of life science image file
formats ( http://www.loci.wisc.edu/software/bio-formats ) .
22.4.2 Data Preprocessing
The raw data coming off of a microscope often suffers from poor contrast, unreg-
istered images, or the lack of precise information about individual structures in the
image data. The following preprocessing steps are commonly employed to prepare
the data for subsequent analysis:
1. Data enhancement: Due to limitations inherent in the microscope and the prepa-
ration of the specimen, the data often suffers from low contrast and high signal-
to-noise ratios. To obtain better results during subsequent steps, the raw data is
improved using image processing filters such as contrast adjustment or equaliza-
tion of lighting. Deconvolution [ 26 ] is commonly used to decrease blurring.
2. Data fusion: Large specimens are often recorded in multiple passes that must be
fused. This is commonly achieved using image stitching [ 45 ] and/or image regis-
tration [ 11 ]. Likewise, multiple recorded channels, e.g., from cameras recording
parts of the sample that fluoresce with different wavelengths, must be registered.
Additional problems arise in time-series data if a living specimen is recorded
that can grow and/or move during recording. Such data not only requires the
registration of individual time-steps, but also the application of drift correction
algorithms.
3. Image segmentation: The image data is commonly too large and complex to be
analyzed manually, thus causing automatic feature extraction algorithms to be
mandatory. In a first step this means data segmentation [ 43 ], i.e., the extrac-
tion of relevant structures, such as cell boundaries, from the data. Segmentation
and object classification is one of the major challenges in computer vision [ 44 ]
and most microscopy images pose additional challenges as the algorithms have
to cope with large variations in data intensity, morphological complexity and
diversity, and varying signal-to-noise ratios. Despite the development of numer-
ous, specialized segmentation algorithms, many applications still require manual
or semi-automatic segmentation.
4. Computation of metadata: Many subsequent analysis steps, especially if time-
series or multimodal data is recorded, require the computation of additional
metadata. Simple examples of metadata include the size and shape of structures,
textural properties, and image statistics of segmented regions. More advanced
metadata is necessary for time series data where numerous cells are captured at
individual time-steps. For the analysis of the morphology and morphodynamics,
the cells need to be tracked over time. Here, cell tracking becomes a crucial part