Datasets and labels¶
The protocol, not the feature directory, defines the task universe. Class order is frozen because it controls integer labels, prompt order, and logit columns.
Registered tasks¶
| Task | Ordered labels | Classes | Current protocol status |
|---|---|---|---|
| TCGA-NSCLC | LUAD, LUSC |
Lung adenocarcinoma; lung squamous cell carcinoma | Registered with frozen metadata universe |
| TCGA-BRCA | IDC, ILC |
Invasive ductal carcinoma; invasive lobular carcinoma | Registered with frozen metadata universe |
| TCGA-RCC | CCRCC, PRCC, CHRCC |
Clear-cell, papillary, and chromophobe RCC | Registered with frozen metadata universe |
| CAMELYON16 | normal, tumor |
Normal lymph node; metastatic lymph node | Registered from official train/test annotations |
| CAMELYON16+17 | normal, tumor |
Normal lymph node; metastatic lymph node | Separate 898-slide FOCUS/MUSE matrix with patient-grouped CAMELYON17 nodes |
| UBC-OCEAN | CC, EC, HGSC, LGSC, MC |
Five ovarian carcinoma subtypes | Registered; local official metadata is required |
The frozen TCGA universes contain 1,043 NSCLC slides from 946 patients, 960 BRCA slides from 900 patients, and 939 RCC slides from 897 patients. CAMELYON16 contains 399 slides in the registered manifest. The separate combined matrix contains 898 slides from 499 patient groups: 399 CAMELYON16 slides plus 499 labeled CAMELYON17 nodes. These counts come from annotations and do not shrink when a feature source is incomplete.
The RCC TITAN store under
/work/hdd/bhwm/master_benchmark/20x_512px_0px_overlap/slide_features_titan
matches all 939 frozen RCC slide IDs. It supplies the 768-wide slide embeddings
used by SLDPC; it is separate from the older BRCA/NSCLC TITAN store under
trident_features.
Metadata contract¶
Every cohort needs columns that resolve to:
- a stable slide identifier;
- a patient or case identifier;
- one label from the ordered task label list;
- optional source-partition or filtering fields.
Protocol compilation rejects blank required metadata fields and slide IDs that
collide after suffix normalization. Split groups carry one strict label by
default. The combined CAMELYON protocol explicitly uses an any tumor group
policy because one CAMELYON17 patient may contain both normal and positive
nodes; few-shot representatives still match the selected target label.
At runtime, split members must all exist in the annotation manifest; loaders do
not silently filter unknown IDs or substitute a root scan when an exact feature
column was configured. CoD-MIL additionally validates every low-to-high patch
map before indexing it.
Patient identity is mandatory for patient-disjoint cancer-cohort folds. CAMELYON16 treats each slide identifier as one case under its current annotation contract. The combined protocol also reports slide-level metrics, but partitions every node from the same CAMELYON17 patient together.
For UBC-OCEAN, place the official training metadata at the path declared by
the protocol. The default additional-task protocol expects columns
image_id, label, and is_tma; TMA composites are excluded. Until the
metadata has rows, generated configs can be valid while metadata_ready,
split_ready, and ready remain false.
Data and feature separation¶
The metadata universe determines who belongs in a fold. A feature registry determines whether a particular method can run for those already-selected slides. Never construct a cohort by listing files in a feature directory: doing so changes the test population for each encoder and invalidates a fair comparison.
Adding a dataset safely¶
Before generating runs, confirm that:
- label aliases have been normalized into one canonical order;
- patient/case IDs cannot leak between partitions;
- exclusions are driven by metadata and recorded explicitly;
- the prompt profile has exactly the same labels and order;
- every feature path template includes a dataset-qualified namespace;
- licensing permits local use of the data and derived features.
See Extending the framework for the registration procedure and Benchmark protocols for the shared fold controls.