Deep MIB - Predict Tab¶

Settings for efficient prediction (inference) and semantic segmentation model generation in Microscopy Image Browser.

How to start the prediction (inference) process¶

Prediction (inference) requires a pretrained network. if you lack one, train it first. pretrained networks can be loaded into Deep MIB for segmenting new datasets.

Steps to start prediction¶

1. Select the pretrained network file in Network filename... in the Network panel. This updates the Train panel with training settings. Alternatively, load a config file via Options tab → Config files → Load>
2. Verify the prediction images directory in Directories and Preprocessing tab → Directory with images for prediction
3. Confirm the output directory in Directories and Preprocessing tab → Directory with resulting images
4. If needed (usually not), preprocess (convert) files:
   • Set Preprocess for: Prediction in the Directories and Preprocessing tab
   • Click Preprocess
5. Switch to the Predict tab and press Predict

Settings section¶

The Settings section configures prediction parameters.

Prediction engine: selects the tiling engine:

• Legacy used until MIB 2.83
• Blocked-image recommended for later versions (supports Dynamic masking and 2D Patch-wise)

Overlapping tiles: (for Padding: same) tiles the patches with overlap to minimize edge artefacts and improve segmentation. Define overlap percentage in %%...

Dynamic masking: skips prediction on certain tiles using on-the-fly masking, configured via the button
Eye previews masking on the current Image View panel image

Padding, %% pads images symmetrically to reduce edge artifacts
Downsample factor for images downsamples prediction images to match training (if applicable), upsampling results to original size and in case of networks with 2 classes are also smoothed
Batch size for prediction... sets the number of patches processed by GPU simultaneously, limited by GPU memory
Model files selects output format for models (CSV for patch-wise)

Score files selects output format for prediction score maps, configured via the button

• upsample predictions: (2D patch-wise only) upsamples downsampled patch-wise predictions to match original image size

Instance segmentation subpanel¶

Settings of the Instance segmentation subpanel are only available for the 2D Instance workflow.

Overlap mode: selects how object instances are stitched across tiles during prediction:

• Centroid in core — each object is emitted by the tile owning its centroid; the tile overlap must exceed the largest object
• IoU merge — detections of neighbouring tiles are merged when their masks agree in the overlap band; works for objects larger than the overlap

An image that is no larger than the network input patch is segmented in one piece, so the overlap mode and the tile overlap settings have no effect on it.

The button configures the stitching parameters:

• Detection confidence threshold (0-1) — minimal confidence score for a detected instance to be kept (both overlap modes); decrease to detect more (weaker) objects, increase to keep only confident detections (default: 0.5)
• Merge IoU threshold (0-1) — (IoU merge only) merge detections of neighbouring tiles when the intersection-over-union of their masks within the shared overlap band exceeds this value; decrease when objects get split at tile seams, increase when distinct touching objects get merged (default: 0.5)
• Merge IoA threshold (0-1) - (IoU merge only) additionally merge when the intersection over the smaller in-band mask area exceeds this value, catching a truncated fragment fully contained in the neighbouring tile's complete mask; applied only to fragments that are really cut by a tile edge (default: 0.8)
• Minimal object area, pixels - (both overlap modes) after stitching, each index is split into its separate objects so that two isolated objects never share one index; components smaller than this are discarded as speckle, set to 0 to keep every component (default: 100)

Merge 2D to 3D stitches the predicted instance models in 3_Results/PredictionImages/ResultsModels into 3D instance models, linking objects that overlap between neighbouring slices into 3D instances with a consistent index. The layout is detected automatically: 2D models are the Z-slices of one stack (taken in alphabetical order as the Z-order) and give a single merged model, while 3D models (predicted from z-stack images) are each stitched independently into one merged model per file. A dialog asks for the stitching settings, then for the destination and file format; the result can be saved as a single 3D file or as a sequence of 2D files. See the 2D Instance workflow for details.

Explore activations¶

The Activations explorer evaluates network performance in detail. It is possible to see details of weights and activation images for all layers and filters.

• Image lists preprocessed prediction images. select one to load a patch matching the network’s input size. use arrows to navigate
• Layer lists network layers. selecting a layer triggers prediction and activation image generation
• Z1..., X1..., Y1... shifts the patch across the image. Update activations with Update
• Patch Z... adjusts Z within 3D network activation patches
• Filter Id... cycles through activation layers
• Update recalculates activations for the current patch
• Collage creates a collage of current layer activations

Preview results section¶

Load images and models loads original images and segmentations into MIB’s active buffer post-prediction

Load models loads segmentations over the current MIB image. Requires that the image is already preloaded.
Load prediction scores loads score images (probabilities) into the active buffer

Evaluate segmentation calculates precision metrics if ground truth labels exist in Labels under PredictionImages. Material names must match training data

Back to MIB | DeepMIB