wsinfer#

Run patch-level classification inference on whole slide images.

wsinfer [OPTIONS] COMMAND [ARGS]...

Options

--backend <backend>#

Backend for loading whole slide images.

Options:: openslide | tiffslide

--log-level <log_level>#

Set the loudness of logging.

Options:: debug | info | warning | error | critical

--version#: Show the version and exit.

run#

Run model inference on a directory of whole slide images.

This command will create a tissue mask of each WSI. Then patch coordinates will be computed. The chosen model will be applied to each patch, and the results will be saved to a CSV in RESULTS_DIR/model-output.

Example:

CUDA_VISIBLE_DEVICES=0 wsinfer run –wsi-dir slides/ –results-dir results –model breast-tumor-resnet34.tcga-brca –batch-size 32 –num-workers 4

To list all available models and weights, use wsinfer-zoo ls.

wsinfer run [OPTIONS]

Options

-i, --wsi-dir <wsi_dir>#: Required Directory containing whole slide images. This directory can only contain whole slide images.

-o, --results-dir <results_dir>#: Required Directory to store results. If directory exists, will skip whole slides for which outputs exist.

-m, --model <model_name>#

Name of the model to use from WSInfer Model Zoo. Mutually exclusive with –config.

Options:: breast-tumor-resnet34.tcga-brca | colorectal-resnet34.penn | colorectal-tiatoolbox-resnet50.kather100k | lung-tumor-resnet34.tcga-luad | lymphnodes-tiatoolbox-resnet50.patchcamelyon | pancancer-lymphocytes-inceptionv4.tcga | pancreas-tumor-preactresnet34.tcga-paad | prostate-tumor-resnet34.tcga-prad

-c, --config <config>#: Path to configuration for the trained model. Use this option if the model weights are not registered in wsinfer. Mutually exclusive with–model

-p, --model-path <model_path>#: Path to the pretrained model. Use only when –config is passed. Mutually exclusive with –model.

-b, --batch-size <batch_size>#

Batch size during model inference. If using multiple GPUs, increase the batch size.

Default:: 32

-n, --num-workers <num_workers>#

Number of workers to use for data loading during model inference (n=0 for single thread). Set this to the number of cores on your machine or lower.

Default:: 2

--speedup, --no-speedup#

JIT-compile the model and apply inference optimizations. This imposes a startup cost but may improve performance overall.

Default:: False

--qupath#

Create a QuPath project containing the inference results

Default:: False

--seg-thumbsize <seg_thumbsize>#: The size of the slide thumbnail (in pixels) used for tissue segmentation. The aspect ratio is preserved, and the longest side will have length max(thumbsize).

--seg-median-filter-size <seg_median_filter_size>#: The kernel size for median filtering. Must be greater than 1 and odd.

--seg-binary-threshold <seg_binary_threshold>#: The threshold for image binarization.

--seg-closing-kernel-size <seg_closing_kernel_size>#: The kernel size for binary closing (morphological operation).

--seg-min-object-size-um2 <seg_min_object_size_um2>#: The minimum size of an object to keep during tissue detection. If a contiguous object is smaller than this area, it replaced with background. The default is 200um x 200um. The units of this argument are microns squared.

--seg-min-hole-size-um2 <seg_min_hole_size_um2>#: The minimum size of a hole to keep as a hole. If a hole is smaller than this area, it is filled with foreground. The default is 190um x 190um. The units of this argument are microns squared.

--patch-overlap-ratio <patch_overlap_ratio>#: The ratio of overlap among patches. The default value of 0 produces non-overlapping patches. A value in (0, 1) will produce overlapping patches. Negative values will add space between patches. A value of -1 would skip every other patch. A value of 0.5 will provide 50%% of overlap between patches. Values must be in (-inf, 1).

tosbu#

Convert model outputs to Stony Brook format.

Convert CSVs of patch predictions to .txt and .json formats for use with Stony Brook Biomedical Informatics viewers.

RESULTS_DIR Path to results directory (containing model-outputs dir).

OUTPUT Path to output directory in which to save files.

wsinfer tosbu [OPTIONS] RESULTS_DIR OUTPUT

Options

--wsi-dir <wsi_dir>#: Required Directory with whole slide images.

--execution-id <execution_id>#: Required Unique execution ID for this run.

--study-id <study_id>#: Required Study ID, like TCGA-BRCA.

--make-color-text, --no-make-color-text#: Make text files with color information for each patch. NOTE: this can add several minutes of processing time per slide.

--num-processes <num_processes>#: Number of processes to use when –make-color-text is enabled.

Arguments

RESULTS_DIR#: Required argument

OUTPUT#: Required argument

patch#

Patch a directory of whole slide iamges.

wsinfer patch [OPTIONS]

Options

-i, --wsi-dir <wsi_dir>#: Required Directory containing whole slide images. This directory can only contain whole slide images.

-o, --results-dir <results_dir>#: Required Directory to store patch results. If directory exists, will skip whole slides for which outputs exist.

--patch-size-px <patch_size_px>#: Required Patch size in pixels.

--patch-spacing-um-px <patch_spacing_um_px>#: Required Physical spacing of the patch in micrometers per pixel.

--seg-thumbsize <seg_thumbsize>#: The size of the slide thumbnail (in pixels) used for tissue segmentation. The aspect ratio is preserved, and the longest side will have length max(thumbsize).

--seg-median-filter-size <seg_median_filter_size>#: The kernel size for median filtering. Must be greater than 1 and odd.

--seg-binary-threshold <seg_binary_threshold>#: The threshold for image binarization.

--seg-closing-kernel-size <seg_closing_kernel_size>#: The kernel size for binary closing (morphological operation).

--seg-min-object-size-um2 <seg_min_object_size_um2>#: The minimum size of an object to keep during tissue detection. If a contiguous object is smaller than this area, it replaced with background. The default is 200um x 200um. The units of this argument are microns squared.

--seg-min-hole-size-um2 <seg_min_hole_size_um2>#: The minimum size of a hole to keep as a hole. If a hole is smaller than this area, it is filled with foreground. The default is 190um x 190um. The units of this argument are microns squared.