Performed a read-only fresh-context review of the uncommitted data_alt.py direct PI-CAI data path and public preprocess_volumes rename in virdx/ve2e.
Compared the draft with the active data.py, VxProsDataModule, VxprosDataset, pipeline table builders and schemas, vxData SDK 1.4.2 materialization, and local PI-CAI schemas.
Confirmed that the direct implementation reproduces canonical NIfTI resample, crop, pad, normalization, transpose, and mask handling and internally removes fake split routing, CSV, config copying, Lightning, DataLoader, and shuffle.
Found that PICAIInference still imports the old path, the returned batch loses patient identity and input ordering, non-RPE ectomy provenance is mislabeled as biopsy, configured dataset upsampling is omitted, and schema-free empty pipeline frames can fail before the intended domain validation.
Reported severity-ranked findings and exact fixes to the parent agent. Made no changes to the ve2e checkout.
Lessons Learned: Pitfalls
A cleaner alternative module does not simplify production behavior until the active model import is replaced; checking call sites is necessary before crediting architectural removal.
preprocess_volumes is the canonical native-volume transform but not the entire VxprosDataset input transform because BaseVxprosDataset.__getitem__ can apply configured in-plane upsampling afterward.
Internally aligned arrays are insufficient for a patient-ID API when pipeline eligibility can drop rows and the output schema carries no identifiers.
The clinical pipeline’s isup_source is a specimen type, not an already-normalized PI-CAI provenance enum; values include RPE, BIOPSY, RESECTION, TURP, and EXCISION.
Lessons Learned: Improvements
Add an inverse SOP defining the complete canonical clinical inference transform, including post-preprocessing upsampling and output geometry.
Define whether vxData sample generation is strict one-requested-patient-to-one-returned-case or a partial eligible-cohort operation; in either case, preserve patient and study identifiers and deterministic order in the returned contract.
Document the allowed mapping from vxData pathology specimen types to PI-CAI label provenance and reject unsupported types rather than coercing them.
Require schemaful empty DataFrames from pipeline steps so downstream table builders can produce stable domain errors for empty cohorts.
