inverse (ve2e)

Goal: accurately infer disease state from bi-parametric MRI using deep inverse models. Filed monthly reports use the tag/name ve2e — treat ve2e and inverse as the same workstream.

Status at a glance (as of the last filed report, March 2026)

• vx1 (DINOv3 ViT encoder, separate classification + segmentation heads) held 1st place on the PI-CAI Open Tuning Leaderboard (0.935 AUROC, 0.774 AP) for patient-level csPCa classification; work has since moved to lesion-level detection/calibration and 4-class ISUP grading.
• FlexDINO: a single model, operating at quadrant/slice/patient field of view via quadrant-masked self-attention + gated attention-based MIL aggregation — built to let simulated (diffsim) data feed training at sub-patient granularity.
• Engineering: the data-platform-based training pipeline is integrated up to preloading (Argo-orchestrated), not yet used for production training; an autonomous LLM-driven experiment-automation stack (logger + structured research context + closed-loop protocol) was stood up, initially targeting late-fusion DINO hyperparameter search.
• SDK migration: per the June 2026 engineering report, ve2e was the first downstream consumer to migrate off vxp_client onto vxdata-sdk 1.0 (2026-06-19) — see vxdata-sdk-and-local-dev.md for the current client API.

Source material

• sources/monthly-reports/ — filed as <MM>-2026_ve2e.md. Jan 2026 is an explicit “no report filed” stub; Feb 2026 has no file at all (unlike diffsim/engineering, no explicit gap marker exists for ve2e Feb — treat as an unknown gap); no report has been filed since March.

No dedicated projects/ or sops/ subdirectories exist yet for this workstream — add them here as inverse/ve2e work gets synthesized from inbox/.