v1.11.0: accept DICOM context (modality/bodyPart/study+series desc) and anchor detection + write-up to it (prevents wrong-organ hallucination)

v1.10.1: FHIR bundle adds ImagingStudy resource + DiagnosticReport.study link when study UID supplied

v1.10.0: MedSAM promptable segmentation /segment, FHIR export /fhir, ICD-10 verification vs NLM clinicaltables

v1.10.0: MedSAM promptable segmentation /segment, FHIR export /fhir, ICD-10 verification vs NLM clinicaltables

v1.9.0: structured per-finding (laterality/location), patient+referrer summaries, TTA prediction-stability confidence, deterministic Fleischner/Lung-RADS/RECIST calculators + /score endpoint

v1.8.3: detector treats radiographic side-markers/view-labels/measurements as non-PHI (no false PHI banner on teaching images)

v1.8.2: remove experimental OOD autoencoder; ICD-10-only coding (no fabricated RadLex/SNOMED); English label normalization; PHI banner only on real identifier concern

v1.8.2: remove experimental OOD autoencoder; ICD-10-only coding (no fabricated RadLex/SNOMED); English label normalization; PHI banner only on real identifier concern

v1.8.1: fix OOD MSE scale (normalize to [-1,1]); PHI red-warn only on likely-identifier

v1.8.0: +skin/retinopathy routing, differentials+scoring+coding, /compare longitudinal, PHI safety check, OOD uncertainty (xrv AE)

v1.8.0: +skin/retinopathy routing, differentials+scoring+coding, /compare longitudinal, PHI safety check, OOD uncertainty (xrv AE)

v1.7.2: pin transformers==4.49.0 (newer ver imported torch.distributed.tensor.device_mesh absent in torch 2.4.1)

v1.7.2: pin transformers==4.49.0 (newer ver imported torch.distributed.tensor.device_mesh absent in torch 2.4.1)

v1.7.1: bone classifier via direct AutoModel (pipelines submodule import was broken)

v1.7.1: bone classifier via direct AutoModel (pipelines submodule import was broken)

v1.7: routed modality (bone-fracture via transformers) + interactive VQA /ask endpoint

v1.6: Grad-CAM heatmap overlay for positive CXR findings (spatial grounding)

v1.5.2: surface segDebug to diagnose empty CTR measurements

v1.5.1: robust CXR gate (radiograph+chest of any view) so frontal CXRs mislabeled 'lateral' still run CNN+CTR

v1.5: anatomy segmentation (PSPNet) -> cardiothoracic ratio measurement; CTR fed to LLM grounding + returned in contract

v1.4: reliable-only LLM grounding (ignore Pneumothorax/Mass/Nodule), default-normal prompt, temperature 0 -> deterministic, stable normal reads

v1.3: VLM image-type detection gates the CXR CNN (no hardcoded chest assumption); CNN runs only on frontal chest radiographs

v1.2: CXR-aware preprocessing (border auto-crop, robust grayscale), 0.5 operating-point positives, debug stats

v1.1: add /analyze (CNN + DashScope qwen-vl-max grounded write-up), warm-load, auth

MedPAC CNN v1: TorchXRayVision DenseNet-121 CXR classifier (GPU FastAPI /infer)

initial commit