---

name: covid-2d-classification description: "Route the 2D COVID chest X-ray and CT classification branch." metadata: disco-role: operating disable-model-invocation: true license: MIT

covid-2d-classification

Use this sub-skill for the 2D COVID classification branch only:

• COVIDx chest X-ray classification
• CovidCT chest CT classification
• COVIDNet / CNN model selection
• train_covid training and validation loops
• MetricTracker, accuracy, and manifest parsing

Do not use this sub-skill for 3D segmentation or 3D COVID segmentation. Route those tasks to the segmentation-oriented or data-preparation sub-skill instead.

Start here

1. Read references/workflows.md for the end-to-end route.
2. Read references/data-layout.md before touching manifests.
3. Read references/api-reference.md for public class and function contracts.
4. Read references/troubleshooting.md if imports or sample runs fail.
5. Run scripts/smoke_covid_imports.py for a safe synthetic import check.

What this branch expects

• 2D RGB inputs only.
• COVIDx for 3-class chest X-ray work.
• COVID_CT for 2-class chest CT work.
• A local MedicalZooPytorch installation, or an equivalent environment where lib.* imports resolve.
• No dependence on the original repository checkout at runtime.

Important caveats

• COVIDxDataset currently has a __getitem__ / load_image keyword mismatch in source.
• CovidNet currently uses pepx instead of PEPX in the constructor.
• CNN uses torchvision backbones with pretrained=True; do not let smoke checks fetch weights.
• MetricTracker averages batch metrics by iteration count, not sample count.

Public entry points

• Dataset loaders: lib.medloaders.COVIDxdataset.COVIDxDataset, lib.medloaders.covid_ct_dataset.CovidCTDataset
• Model factory: lib.medzoo.create_model
• Training loop: lib.train.train_covid.train, lib.train.train_covid.validation
• Metrics: lib.utils.covid_utils.accuracy, lib.utils.covid_utils.MetricTracker
• Helper parser: lib.utils.covid_utils.read_txt