KPMP-FOCAL: tissue masks and regional review records

This package contains existing FOCAL outputs for all 181 processed KPMP H&E slides. It is prepared for review by the KPMP Data Coordinating Center. It includes tissue masks, image previews, regional decisions, and prior visual-review records. The release includes all processed slides, not only the final quantitative evaluation subset.

FOCAL stands for Foundation model and Oracle Classification for Annotation-free tissue Localization. These outputs are offered for evaluation and reuse.

Example comparison

Slide 30-11120 illustrates a visible tissue area retained by FOCAL but absent from the spatial reference. The gold outline marks the reference's valid area. The four-example PDF provides additional comparisons.

Start here

Open slide_manifest.csv to find a slide and its files. The paths are relative to this folder. View previews/comparison/<slide_id>.jpg for the original image, FOCAL overlay, and spatial-reference overlay side by side. Green indicates tissue. The gold boundary marks the reference's valid area. Regions outside that boundary are unlabelled by the reference.

Files

Location	Contents
slide_manifest.csv	One row per slide, file locations, dimensions, and review status
masks/	181 original FOCAL Native masks on the working image grid
previews/original/	Existing colour thumbnails, copied without alteration
previews/focal_overlay/	FOCAL masks over the original thumbnails
previews/reference_overlay/	Spatial-reference masks and their valid-area boundaries
previews/comparison/	Three-panel previews for all 181 slides
annotations/region_annotations.csv	1,604 recorded regional decisions, labels, and explanations
annotations/slide_annotation_summary.csv	Region counts and mask areas for each slide
annotations/original_records/	Original decision files, region geometry, spatial inputs, and provenance
reference_review/reference_review.csv	Existing review flags, example observations, and quantitative-selection history
reference_review/spatial_masks/	Existing spatial-reference masks
reference_review/valid_regions/	Masks identifying where the spatial reference is defined
reference_review/examples/	The existing four-example publication panel and individual comparisons

Interpreting the masks

FOCAL masks use zero for background and nonzero for tissue. Masks and thumbnails share the working grid. Original WSI dimensions and separate horizontal and vertical scale factors are listed in the manifest. Original full-resolution WSIs are not included.

For a working pixel centre (x, y), the corresponding source-image centre is ((x + 0.5) * scale_x - 0.5, (y + 0.5) * scale_y - 0.5). This mapping locates an existing working-grid mask on the source image. It does not add full-resolution boundary detail.

The spatial references were derived from released Space Ranger spot-level in_tissue labels. Pixels inside the convex hull of all spot centres inherit the nearest spot's label.

Existing alignment corrections for 28-12265-XY02 and 28-12265-XY03 are retained. Their downloaded spatial files cross-match the other image. The manifest records the corrected spatial-source participant. The supporting record is in annotations/original_records/cohort_records/alignment_corrections.json.

Regional annotations

Regional decisions and artifact labels were generated by FOCAL. They have not all been independently verified by an expert. reason retains the recorded explanation. confidence is the model's categorical assessment, not a calibrated probability. native_decision and outline_decision preserve the two recorded mask conventions.

Coordinates in region_annotations.csv are zero-based working-image coordinates, with the origin at the upper left. Bounding boxes include the minimum coordinates and exclude the maximum coordinates. They bound a region but are not artifact contours. The losslessly compressed label_map.npy.gz identifies region pixels through label_map_value. Zero denotes pixels outside the explicitly listed regions. The original base mask and disagreement map are also retained.

Label names are preserved from the original files. Examples include pen_ink, debris_fibre, coverslip_mountant, and out_of_focus. bare_glass and boundary_halo describe background or boundary context. not_applicable means no artifact type was assigned. Region counts are counts of recorded decisions, not independently counted physical artifacts. The package does not contain a separately validated artifact-segmentation mask.

Existing visual review

Fifty slides were explicitly flagged in the original manual review. Detailed error subtypes were not recorded for most of them. The review CSV preserves the original comments.

The four examples are 30-10868 (an agreement example), 29-10423 (fragmented tissue), 30-11120 (a large omitted tissue area), and 31-10100 (darker, lower-contrast tissue).

Integrity and provenance

checksums.sha256 covers all package files except itself. Run sha256sum -c checksums.sha256 from this folder to check them. validation.json records the packaging checks. Original input locations and SHA-256 hashes are retained in annotations/original_records/file_provenance.csv. Existing input masks and thumbnails were verified against the frozen 181-slide review index. JPEG overlays and comparisons are new viewing aids. They are not analytical masks.