New Features

Deep Learning & AI

New features for loading a model

The following features have been added to the ReadOnnxModel algorithm:

  • A new forceCpu option allows preventing the model from being loaded into GPU memory. This parameter will become mandatory in C++ starting from version 2027.1.
  • Model paths can now be specified relative to an environment variable using the ${VARIABLE_NAME} syntax.

Support for new input layouts

Metadata-based inference

The new OnnxMetadataBasedPrediction module enables deep learning inference driven by image metadata.

This feature allows dynamic adaptation of ONNX model predictions based on metadata associated with the input image, enabling more flexible and context-aware AI workflows.

It is particularly useful when working with models trained using Avizo, as metadata stored with the model can be directly leveraged without additional configuration.

Segmentation

Island removal

The RemoveIslands algorithm removes connected regions smaller than a user-defined size and merges them with neighboring regions based on configurable criteria.

This is particularly useful for cleaning segmentation results.

Segmented object extraction

The ExtractSegmentedObjects algorithm extracts individual objects from a label image.

Each labeled region is isolated, enabling object-level analysis, classification, and reporting.

Image Editing

Resizing preserving ratio

The ResizeWithRatio2d and ResizeWithRatio3d algorithms adapt the resolution of an image while preserving its aspect ratio.

If scaling factors differ between axes, padding is applied using a user-defined value.

Application of a colormap

The LabelToColor and BinaryToColor algorithms convert grayscale label images into RGB or RGBA images by mapping label values to colors.

The resulting images can be combined with the original image using the BlendAlphaChannelWithImage algorithm.

Drawing utilities for 2D images

Two new algorithms are introduced for rendering rectangular primitives on 2D images:

These tools are particularly useful for annotation and visualization, such as displaying bounding boxes from detection algorithms.

Drawing of rectangles around segmented channels of a 3D NAND

Image Filtering

Spatial FFT Filter

The FftSpatialFilter2d and FftSpatialFilter3d algorithms filter an image using its FFT magnitude.

They are particularly useful for removing or isolating periodic noise in images.

Deconvolution algorithm with low memory footprint

The RichardsonLucyDeconvolution3d algorithm provides a tiled 3D implementation of the Richardson–Lucy deconvolution, significantly reducing memory usage by processing sub-volumes instead of the full image.

Two tiling modes are available:

  • USER_DEFINED: reduces memory usage while preserving identical results
  • USER_DEFINED_LOW_MEMORY: further reduces memory usage with possible minor differences in results

Background estimation

The BackgroundImageFromDownsampling2d and BackgroundImageFromDownsampling3d algorithms estimate image background using bicubic downsampling.

They are designed to be used with ShadingCorrection to improve signal uniformity and enhance feature visibility.

Color standardization

The new ColorStandardization operator has been introduced to normalize and standardize grayscale or color images. The resulting color image has channels typically between -1 and 1.

Default parameters follow the PyTorch ImageNet preprocessing convention, enabling direct integration into common machine learning workflows.

Quantification and Analysis

Segmentation Metrics improvement

The SegmentationMetrics algorithm now supports a macro-averaging mode, in addition to the existing micro-averaging.

This mode computes metrics independently for each class and averages the results, which is better suited for many deep learning use cases.

The output data frame has also been improved for better consistency.

Similarity metric extension

The SimilarityMetricValue algorithm now supports a new NORMALIZED_MUTUAL_INFORMATION mode.

This metric is particularly useful for multi-modal image comparison where intensity relationships are non-linear.

Statistical analysis

The new AnalysisStatistics algorithm computes statistical measurements from tabular data.

It enables efficient analysis and reporting of numerical results.

Enhancements

Performance improvements

The following features are faster to execute:

Compatibility changes

API removal

  • ConvolutionWithImage2d was deprecated since 2024.3 and has been removed in 2026.1. The ConvolutionWithImage algorithm can be used instead.
  • FlipImage2dand FlipImage3d were deprecated since 2024.3 and have been removed in 2026.1. The Flip2d and Flip3d algorithms can be used instead.
  • RescaleImageByFactor was deprecated since 2024.3 and has been removed in 2026.1. The RescaleByFactor algorithm can be used instead.
  • GradientOperator2d was deprecated since 2023.1 and has been removed in 2025.2. The GradientVector2d and GradientMagnitude2d algorithm can be used instead. The CANNY_DERICHE, SHEN_CASTAN, CANNY, and AMPLITUDE_AND_ORIENTATION modes have no equivalent in the new API.
  • GradientOperator3d was deprecated from 2023.1 and has been removed in 2025.2. The GradientVector3d and GradientMagnitude3d algorithm can be used instead. The CANNY_DERICHE, and CANNY modes have no equivalent in the new API.
  • The Image Data helper functions were deprecated from 2023.2. The equivalent methods are available as members of the ImageView class of IOLink.
    • ImageDev getImageInterpretation and setImageInterpretation helper methods are replaced by imageInterpretation and setImageInterpretation IOLink ImageView methods.
    • ImageDev getDimensionalInterpretation and setDimensionalInterpretation helper methods are replaced by dimensionalInterpretation and setDimensionalInterpretation IOLink ImageView methods.
    • ImageDev getCalibrationOrigin and setCalibrationOrigin helper methods are replaced by calibrationOrigin and ImageDev setCalibrationOrigin IOLink ImageView methods.
    • ImageDev getCalibrationSpacing and setCalibrationSpacing helper methods are replaced by calibrationSpacing and ImageDev setCalibrationSpacing IOLink ImageView methods.
    • ImageDev getCalibrationUnit and setCalibrationUnit helper methods are replaced by calibrationUnit and setCalibrationUnit IOLink ImageView methods.

Breaking changes

  • The API of the SegmentationMetrics algorithm has changed
    • The PER_IMAGE value of the metricScope enumerate parameter has been renamed MICRO_AVERAGED.
    • The MACRO_AVERAGED value has been added to the metricScope enumerate parameter.
    • The outputImageMeasurement output has been renamed
    • The outputMacroMeasurement output has been added.
  • The LoadOnnxModel algorithm has a new parameter

Platform support changes

ImageDev is no longer available for Visual Studio 2017 (VC 15).

  • ONNX prediction features for GPU execution are now compatible with cuDNN versions ≥ 9.10. Previous versions required cuDNN 8.4.2 and were not compatible with cuDNN 9.x..

Operating systems

ImageDev 2026.1 is available for:

  • Windows 64-bit: Visual Studio 2019 (VC 16), and 2022 (VC 17).
  • Linux 64-bit: GCC 9 (Ubuntu 20.04), and GCC 11 (Ubuntu 22.04).

Deprecations and future removals

ImageDev 2026.2

  • The GCC 9 packages for Ubuntu 20.04 will be no longer provided from ImageDev 2026.2.
  • The VC16 packages for Visual 2019 on Windows will be no longer provided from ImageDev 2026.2.

ImageDev 2027.1

  • This forceCpu parameter will be mandatory in the readOnnxModel C++ function from the 2027.1 version