# PDR 006: Match pathology model development to each task's supervision

- HTML version: https://robbiepalmer.me/projects/pathologize/pdrs/006-task-native-pathology-model-development
- Project: Pathologize
- Status: Proposed
- Opened: 2026-11-07

## Context

[PDR 003](/projects/pathologize/pdrs/003-independent-pathology-benchmark-families)
separates quality, tiling, tissue, cell, representation, and slide-level
benchmarks. Those families still contain different learning problems. Patch
classification, continuous regression, object detection, semantic and instance
segmentation, and slide-level multiple instance learning do not share one
label shape or useful error measure.

Forcing every problem into image classification would discard location,
boundaries, counts, continuous values, or the distinction between instance and
bag supervision. Forcing them through one model framework would make the
framework's supported tasks define the research programme.

## Proposed decision

Match each benchmark question to the supervision and prediction type that
expresses it directly. Pathologize will support:

* single-label, multilabel, and ordinal image or patch classification;
* continuous image, region, or slide regression;
* object and nucleus detection;
* semantic, instance, and panoptic segmentation;
* multiple instance learning over variable-size bags of patch features, with
  optional coordinates and more than one slide per subject; and
* later task types when they add a typed prediction contract, leakage policy,
  baseline, metrics, resource limits, and inspectable failures.

Keep these as task-specific benchmark contracts. A common run envelope may
record model identity, inputs, resources, failures, and artifacts, but it must
not erase the task's native predictions. A slide-level result may consume a
pinned patch encoder or cell model without turning either upstream output into
ground truth.

Compare models through Pathologize metrics and fixed project splits. Framework
metrics and default dataset splits are diagnostic only. Do not require one
training library to support every task.

Choose model and training approaches by development-set quality, robustness,
compute cost, iteration time, and implementation burden. Keep final test and
external cohorts sealed. A published paper's setup is one candidate among
several and has no privileged status.

## Consequences

Pathologize can choose a productive library for each task without changing the
meaning of a benchmark. Classification and regression can use transfer
learning, dense tasks can preserve spatial outputs, and weak slide labels can
remain bags rather than fabricated patch labels.

The project must maintain several prediction schemas, metrics, and
framework-specific experiment entrypoints.
Cross-task comparisons need explicit dependency versions and cannot collapse
into one leaderboard score. Some models will produce several useful outputs
and need evaluation in more than one contract.

## Evaluation

Accept the proposal when library-native experiments produce one image or patch
prediction baseline, one spatial detection or segmentation baseline, and one
slide-level multiple instance baseline from pinned Pathologize splits.
Each run must publish its native prediction records, resource measurements,
failure slices, and model provenance. Reject the proposal if adding a framework
requires its dataset format, split policy, or native score to become the
benchmark contract.

## Related records

- Project: [Pathologize](https://robbiepalmer.me/projects/pathologize.md)
- Implemented by ADR: [ADR 004: Framework-native training with common result contracts](https://robbiepalmer.me/projects/pathologize/adrs/004-framework-native-training-result-contracts.md)
- Implemented by ADR: [ADR 005: fastai and timm transfer-learning baselines](https://robbiepalmer.me/projects/pathologize/adrs/005-fastai-timm-transfer-learning-baselines.md)
- Implemented by ADR: [ADR 006: Ultralytics detection and instance-segmentation baselines](https://robbiepalmer.me/projects/pathologize/adrs/006-ultralytics-spatial-instance-baselines.md)
- Implemented by ADR: [ADR 007: Transformers foundation-model workflows](https://robbiepalmer.me/projects/pathologize/adrs/007-transformers-foundation-model-workflows.md)
- Implemented by ADR: [ADR 008: torchmil multiple-instance-learning baselines](https://robbiepalmer.me/projects/pathologize/adrs/008-torchmil-multiple-instance-baselines.md)
- Evidence: [Patho-Bench public pathology tasks and evaluation methods](https://github.com/mahmoodlab/patho-bench)
- Evidence: [Attention-based deep multiple instance learning](https://proceedings.mlr.press/v80/ilse18a.html)
- Evidence: [Panoptic quality analysis for computational pathology](https://pmc.ncbi.nlm.nih.gov/articles/PMC10224986/)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt