CLI entrypoint for batch inference

[miguel5andrade]

Hi!

Is it worth having a CLI entry point to run inference on multiple WSIs? classpose-predict-wsi only allows inference on one WSI at a time... while it's easy for a user to write a script to loop through all WSIs in a directory and call predict_wsi.py, it would be better to have it out of the box.

Maybe create a new CLI entry point or have classpose-predict-wsi accept a --slide_dir_path argument?

[josegcpa]

Hi @miguel5andrade! We are currently not supporting this directly as people tend to get very creative with how they store slides 👀 I think the easier way to adapt this would be with something like

import argparse
import os
from pathlib import Path
from copy import deepcopy
from classpose.entrypoints.predict_wsi import main, DEFAULT_TILE_SIZE, DEFAULT_OVERLAP

def main_with_args():
    parser = argparse.ArgumentParser(description="Run Classpose WSI batch inference.")

    parser.add_argument(
        "--model_config",
        type=str,
        required=True,
        help="One of 'conic', 'consep', 'glysac', 'monusac', 'puma' or "
        "a path to the Classpose model config.",
    )
    parser.add_argument(
        "--slide_dir_path",
        type=str,
        required=True,
        help="Path to directory containing WSI files.",
    )
    parser.add_argument(
        "--tissue_detection_model_path",
        type=str,
        default=None,
        help="Path to the GrandQC tissue detection model weights. If specified but not found"
        " it will be downloaded.",
    )
    parser.add_argument(
        "--artefact_detection_model_path",
        type=str,
        default=None,
        help="Path to GrandQC artefact detection model. If specified, detects artefact regions.",
    )
    parser.add_argument(
        "--filter_artefacts",
        action=argparse.BooleanOptionalAction,
        default=False,
        help="If True, filters cells detected in artefact regions. Requires --artefact_detection_model_path.",
    )
    parser.add_argument(
        "--roi_geojson",
        type=str,
        default=None,
        help="FeatureCollection with (Multi)Polygon(s) in level-0 coordinates",
    )
    parser.add_argument(
        "--roi_class_priority",
        type=str,
        default=None,
        nargs="+",
        help="Space-separated list of ROI class names in priority order for overlapping regions "
        "(used for density calculations only). "
        "Cells in overlapping ROIs will be assigned to the first matching class in this list. "
        "Classes not in this list will be checked after, in their natural order. "
        "Example: Tumour Stroma Necrosis",
    )
    parser.add_argument(
        "--min_area",
        type=int,
        default=0,
        help="Minimum area of the tissue polygons.",
    )
    parser.add_argument(
        "--tta",
        action=argparse.BooleanOptionalAction,
        default=False,
        help="Triggers test time augmentation.",
    )
    parser.add_argument(
        "--batch_size",
        type=int,
        default=8,
        help="Batch size for inference.",
    )
    parser.add_argument(
        "--device",
        type=str,
        default=None,
        help="Device to use for inference. If None, automatically infers the device. "
        "Multi-GPU execution can be specified as 'cuda:0,1' or 'cuda:0,1,2,3'.",
    )
    parser.add_argument(
        "--bf16",
        action=argparse.BooleanOptionalAction,
        default=False,
        help="Triggers bf16 inference.",
    )
    parser.add_argument(
        "--tile_size",
        type=int,
        default=DEFAULT_TILE_SIZE,
        help="Tile size for inference.",
    )
    parser.add_argument(
        "--overlap",
        type=int,
        default=DEFAULT_OVERLAP,
        help="Tile overlap for inference.",
    )
    parser.add_argument(
        "--output_folder",
        type=str,
        help=(
            "Path to save the folder containing the folders with output files "
            "(basename_cell_contours.geojson, basename_cell_centroids.geojson, "
            "basename_tissue_contours.geojson, basename_artefact_contours.geojson)."
        ),
        required=True,
    )
    parser.add_argument(
        "--output_type",
        type=str,
        default=None,
        nargs="+",
        choices=["csv", "spatialdata"],
        help="Optional output type(s). Accepts one or more of: 'csv', 'spatialdata'. "
        "'csv' generates cellular density statistics (cells/mm²) per class. "
        "'spatialdata' generates a unified SpatialData Zarr store containing all outputs. "
        "Can specify multiple types separated by spaces (e.g., --output_type csv spatialdata).",
    )
    args = parser.parse_args()

    output_path = args.output_folder # this will be replaced
    for slide in Path(args.slide_dir_path).glob("*"):
        if slide.suffix in ["svs", "tif", "tiff", "ndpi"]: # minimal list of suffixes to avoid reading the wrong files
            curr_args = deepcopy(args)
            output_path = os.path.join(output_path, str(slide).replace(slide.suffix, ""))
            curr_args.slide_path = str(slide)
            curr_args.output_path = output_path
            main(curr_args)

if __name__ == "__main__":
    main_with_args()

This is not optimised though! The model gets loaded multiple times, etc. It's not ideal, but this provides a very simple wrapper to do just that. We will try to consider this in the future, it's not an uncommon application but we just haven't gotten to it :-)

[miguel5andrade]

yes that makes sense! thank you!