Formation, processing, analysis, and display of images and video.
Continued V-JEPA-2.1 pretraining on approximately 2,650 hours of surgical video improves partially fine-tuned performance across six robotic-surgery task families.
Radiologist-reviewed masks raise measured nnU-Net Dice scores by 0.143–0.188, exceeding score changes from altering training-data composition.
A 60 Hz eye-tracking dataset links radiologists’ decision windows to lesions, raising a 3D nnU-Net baseline from 0.6008 to 0.6819 Dice.
Auditing report-derived MIMIC-CXR labels against radiologist image review found only 1% case capture; a curated-cohort DenseNet121 reached ROC-AUC 0.853.
Paired image and report annotations expose label-source bias, including 59% of image-positive findings absent from reports.
HistoFID normalizes each distance to an encoder-specific cohort floor, cutting cross-encoder variation by 89% within cohort.
A compression-trained pixel diffusion codec gates misaligned DMD updates, cutting bitrate by 62.5% at matched LPIPS versus GLVC.
Group-of-Latents sends sparse anchor latents plus text, then uses a DiT denoiser to synthesize missing video latents at no extra bitrate.
FoundCAC pretrains on a stratified lens library and retrieves vector-quantized PSF features, reaching 26.84 dB average PSNR on RealLens-Sim.
Sobel-gated boundaries, multi-directional Mamba context, and skeleton-guided fusion raise junior oncologist IoU from 0.899 to 0.965 and cut contouring time by over 80%.
