clam
49Skill Claude CodeCodex
De-facto modern open-source weakly-supervised MIL aggregator. Per-class attention branches + instance-level clustering loss.
Skill Claude CodeCodex
De-facto modern open-source weakly-supervised MIL aggregator. Per-class attention branches + instance-level clustering loss.
Skill Claude CodeCodex
Canonical CNN for joint nuclei instance segmentation + classification on H&E. Three-decoder architecture with horizontal-vertical distance maps for instance separation.
Skill Claude CodeCodex
Pretrain a patch encoder on millions of unlabelled pathology patches using SSL (contrastive / DINO / MIM). Frozen features feed downstream MIL aggregators.
Skill Claude CodeCodex
Umbrella method for any large pretrained pathology encoder. Use this node to position a new model / paper / benchmark in the PFM space.
Skill Claude CodeCodex
Umbrella method for pathology vision-language models. Aligns image features with text via contrastive (CLIP-style) or generative pretraining; enables zero-shot classification, retrieval, captioning, and VQA.
Skill Claude CodeCodex
Transformer-based MIL aggregator with Nyström attention + PPEG positional encoding. Models patch-patch correlations at gigapixel scale.
Skill Claude CodeCodex
Umbrella method for weakly-supervised multiple instance learning on WSIs. Train a slide-level classifier from slide-level labels alone.
Skill Claude CodeCodex
Swin-T pathology SSL checkpoint, pretrained on TCGA+PAIP with SRCL (a MoCo v3 variant). The pre-UNI open-source pathology SSL baseline.
Skill Claude CodeCodex
CLIP ViT-B/32 fine-tuned on 200k pathology image-caption pairs from medical Twitter (OpenPath). The first widely-cited open-weights pathology VL model.
Skill Claude CodeCodex
Tell Claude when to recommend UNI as a backbone, how to load it, and where it fails.
Skill Claude CodeCodex
Define a benchmark's goal, datasets, tasks, metrics, baselines, and the workflows for adding models / datasets / aggregating results.