Quantitative image analysis of microscopy datasets is a central tool in modern cell and developmental biology (Meijering and Cappellen, 2007; Senft et al., 2023). The advent of machine learning methods has established Python as the de facto standard for the analysis of biomedical images (Jacquemet, 2021; Morgado et al., 2024). Jupyter notebooks (Granger and Pérez, 2021) provide a beginner-friendly environment that enables prototyping and distribution of complex image processing pipelines (von Chamier et al., 2021). Platforms such as Google Colab or DigitalOcean provide free access to high-end computing capacity to execute notebooks, thus "democratizing" the use of resources that would otherwise be too costly. Notebooks facilitate data exploration and increase the reproducibility of data processing pipelines (Kluyver et al., 2016; Pimentel et al., 2019; Samuel and Mietchen, 2024). The leading position of notebooks for data analysis is cemented by the recent availability of large language models within notebooks (Qiu, 2023), which further facilitates development, debugging and execution of image processing and quantification routines. Thus, notebooks are a central element in the analysis of biomedical image data.

Both biological and medical imaging generate multidimensional data. Confocal and light sheet microscopies, for instance, allow collection of three-dimensional stacks (X, Y, Z) of different fluorescent reporters (wavelength), over the duration of a biological process (time). Multimodality imaging, such as correlative confocal and electron microscopies (Casares-Arias et al., 2021), confocal and super-resolution (Xiang et al., 2018), or the combination of anatomical and functional data in medical imaging (Pichler et al., 2008), introduces a sixth imaging dimension. Processing and analysis of multidimensional image datasets often requires image exploration at many stages in which rapid visualization and browsing are key steps. Two-dimensional images can be easily displayed in notebooks, e.g. with matplotlib (Hunter, 2007). However, few tools exist to display, browse and interact with multidimensional images in a notebook (Haase, 2021), and some of the available options provide complex user interfaces and functionality that prevent light-weight, inline multidimensional image browsing (Sofroniew et al., 2025).

Microscopy images are generally archived in remote backup servers. The advent of cloud-based storage services has only increased this trend. Remote and cloud-based servers often do not provide a graphical user interface, but simply a terminal-based, command line interface. Terminals can display text-based information about the images (file name, size, date, etc.), but do not allow image visualization. The lack of tools for image display on the terminal makes it hard to select images either for further processing or to generate figures for publication.

Here we introduce JuNkIE-CLImax, a pair of open source tools (https://bitbucket.org/rfg_lab/junkie and https://bitbucket.org/rfg_lab/climax) for the display and interactive visualization of multidimensional microscopy images in environments traditionally considered to be document or text-based. JuNkIE is a JUpyter NotebooK Image Explorer that enables the opening, display and simple manipulation of multidimensional images in Jupyter notebooks. CLImax, a Command Line IMAge eXplorer, allows users to open, visualize and browse multidimensional images on the terminal, using character-based strategies for image rendering.

JuNkIE-CLImax: multidimensional image explorers

JuNkIE-CLImax can load multidimensional images in the multi-page TIFF format often found in microscopy (Besson et al., 2019), as well as images stored in the increasingly popular zarr array format (Miles, 2015). JuNkIE-CLImax can also import file sequences that represent different dimensions ( e.g. time or wavelength) of a multidimensional image. Internally, images are stored using numpy , a package for numerical computation, and its ndarray data structure to represent multidimensional arrays (Harris et al., 2020). JuNkIE-CLImax display a two-dimensional view of a multidimensional image ( Fig. 1A ). Both packages provide sliders to select a specific Z-plane, time point and channel from the multidimensional image stored in memory, using both mouse and keyboard-based interactions. JuNkIE-CLImax continuously update the displayed image in real time in response to user input. To enable use in settings with reduced compute power ( e.g. remote storage servers), it is possible to disable the continuous update of the display, such that updates only occur when the user interaction is completed ( e.g. when the mouse button is released, but not while the mouse is dragged). XY, XZ or YZ sectioning planes can be selected. Additionally, JuNkIE-CLImax provide functionality for simple image manipulation. This includes rotation and flipping (both horizontal and vertical), colormap selection and inversion (including all the colormaps available in matplotlib (Crameri et al., 2020; Crameri, 2023)), and both manual and automated contrast adjustment.

JuNkIE

JuNkIE displays images using the Figure class in matplotlib . A matplotlib Figure can hold different plot elements, including images. The matplotlib Figure provides a toolbar with options to pan, zoom in and out, reset the view, or save a snapshot ( Fig. 1A, top).

JuNkIE provides a simple application programming interface (API). Using the JuNkIE API it is possible to open images from disk, by providing a file or folder path, or to view volumetric information already loaded as an ndarray . Substrings can be used to distinguish which files in an image sequence store specific dimensions ( e.g. different image channels). Lastly, the colormap used to initially display an image, as well as the size of the figure can be determined with input parameters when invoking JuNkIE.    

JuNkIE is distributed with a test suite using pytest (Krekel et al., 2004) and the nbval plugin for notebook-based testing (Cortes-Ortuno et al., 2016), with code coverage over 90%.

CLImax

To display images on the terminal, we created a character-based image renderer. The image renderer uses rich , a Python package for text formatting that allows the use of coloured characters in the terminal (McGugan, 2019; Burns, 2022). Our image renderer takes a two-dimensional ndarray as its input, and produces a sequence of rich Segments, each consisting of a character and some formatting (a Style ), punctuated by end-of-line characters at the end of each image row ( Fig. 1B ). rich Styles determine the colours used to display a character and its background. Colours are defined by the corresponding red, green and blue (rgb) components, specified as a string. Initially, we rendered each pixel using a space character (' '), assigning the pixel value as the background colour and no foreground colour. Because characters on the terminal are rectangular (elongated along the vertical axis), we rendered each image column twice to obtain square pixels. As a consequence, image rendering was relatively slow, requiring 380±23 ms (mean±standard deviation) to render a 512x672 pixel image on an Apple M3 Max equipped with 48 GB of RAM.

To optimize image rendering, we first reduced the number of terminal characters necessary to display a pixel. We took advantage of the half-block character ('▄') to visualize pairs of vertically-adjacent pixels (same column, consecutive rows) with a single terminal character ( Fig. 1B ). For each pair of pixels, we rendered the bottom pixel by assigning its value as the foreground colour of a half-block character, and the top pixel by assigning its value as the background colour of the same half-block character ( Fig. 1B ). This approach allowed us to display each pixel with half a character, in contrast to the previous method, which required two characters per pixel. We further accelerated rendering by creating a Style cache that used a Python dictionary to store the rich Styles corresponding to parsed colour-strings. After these optimizations, rendering speed increased by 57% (163±13 ms/image) and memory consumption decreased by a factor of four.

To further speed up image display, we implemented the hot-loop of the renderer in Rust (Matsakis and Klock II, 2014). The Rust code implements a helper function with two inputs, a two-dimensional image, and a colour lookup table that contains the string representation of each of the colours available in the current colormap. The helper function then scans the image columns, two rows at a time, generating the Style string for the character that will display a pair of vertically-adjacent pixels. We parallelized the scan over the image columns using the Rust crate rayon (Matsakis and Stone, 2014) to create a parallel iterator. Overall, Rust-based image rendering was 11% faster (146±13 ms/image) than our fastest Python implementation.

We designed the CLImax user interface with textual (McGugan, 2021) ( Fig. 1A, bottom). textual is a Python package for the development of user interfaces and applications for the terminal. textual provides different widgets ( e.g. buttons, dropdown boxes, text labels, etc.). We created a new ImagePanel widget that displays an image using the character-based renderer described above. textual apps provide a command palette that includes options to change the theme used to display the application, save a screenshot, or show a help panel. In addition, CLImax provides options to open a new image, zoom in or out, display the image with bilinear interpolation (particularly useful when zooming in/out), or save the current display.

CLImax can be invoked from the command line with different options. Similar to the JuNkIE API, the CLImax command line parameters enable image selection and substring and colour map specification, as well as initial zoom level, particularly important given that image rendering in the terminal is computationally intensive, with a cost that increases with the number of pixels to be rendered.

CLImax is also tested with pytest , using the asyncio plugin (Seifert, 2015) to test asynchronous user interaction. CLImax tests provide code coverage greater than 90%.