Performed a read-only architecture investigation of virdx/ve2e branch feat/isup-inference.
Traced src/ve2e/inference_api/picai/data.py through its 24-step vxData pipeline, materialization, Lightning data module, clinical dataset, and canonical resample/crop/pad/z-score preprocessing path.
Searched source, scripts, tests, and configs for inference, artifact loading, dataset dumping, materialization, and preprocessing seams.
Found that model input preparation, benchmark cohort selection, pathology target construction, expert-mask selection, split routing, and training-table construction are mixed in one 254-line module.
Recommended keeping PICAIInference array-only, moving benchmark cohort and target creation out of the general inference API, freezing native-data benchmark manifests as immutable vxData artifacts, and loading them through the existing materialization and clinical dataset preprocessing path.
Identified the current shared canonical preprocessing seam in clinical_dataset._preprocess_vols, which is already used by both eager and on-demand clinical dataset loading.
Made no changes to the ve2e checkout.
Lessons Learned: Pitfalls
The initial workstream inference was too broad. The repository README calls ve2e “the inverse model”, so the inverse workstream index should be read before product or infrastructure context.
Repository-wide searches can be noisy because this branch includes a large restructuring. Narrow symbol and AST-based traces were more useful than branch-wide diffs.
The current inference-data module is not covered by tests and the sole benchmark caller passes the sample-loader arguments in the opposite positional order. Static API tracing is therefore important before assuming the branch path has run successfully.
Lessons Learned: Improvements
The inverse workstream would benefit from a short SOP that records the canonical native-MRI preprocessing boundary and distinguishes immutable training/benchmark artifacts from online inference inputs.
A documented benchmark-manifest contract should define cohort immutability, expert-mask provenance, target provenance, and allowed split semantics without overloading the training table builder.
The public inference API documentation should state that PICAIInference consumes model-ready arrays only and that data acquisition and benchmark targets belong to separate adapters.
