Clarify checkpoint selection and Table 1 scope

README.md

@@ -14,8 +14,8 @@ tags:
 
 # TiSage Segmentation Checkpoints
 
-Paper checkpoints for **TiSage: Tissue Segmentation with Multi-Scale Semantic
-Guidance**, released with the code and reproducibility material at
+Trained checkpoints associated with the paper settings for **TiSage: Tissue
+Segmentation with Multi-Scale Semantic Guidance**, released with the code and reproducibility material at
 [carlosh93/TiSage](https://github.com/carlosh93/TiSage). The paper was selected
 as a **Spotlight** at the Eleventh ISIC Skin Image Analysis Workshop @ MICCAI
 2026.
@@ -30,9 +30,11 @@ as a **Spotlight** at the Eleventh ISIC Skin Image Analysis Workshop @ MICCAI
 | `dfutissue_unimatch_v2_fixed_seed1_best.pth` | DFUTissue | UniMatch-V2 | fixed/full supervision split, seed 1 |
 
 Each file contains `model`, `model_ema`, optimizer state, epoch, and validation
-selection metadata. TiSage evaluation defaults to `model_ema`; pass an explicit
-state to the evaluator when comparing student and teacher weights. The LUTSeg
-pair is also used by the repository's Figure 4 reproduction script.
+selection metadata. The files are best-student snapshots and contain the EMA
+state from the same epoch. The paper's Table 1 numbers are selected best-EMA
+values across training and remain traceable to the code repository's committed
+logs, so an individual released snapshot can differ from the Table 1 value.
+The LUTSeg pair is also used by the Figure 4 reproduction script.
 
 ## Download
 
@@ -75,7 +77,9 @@ python method/eval/evaluate_checkpoint.py \
 The public LUTSeg release has a patient-disjoint validation set and no separate
 public test split. The evaluator reports per-class IoU and Dice plus their
 means on the 30-image validation set and requires one CUDA GPU. Append
-`--check-only` to validate checkpoint compatibility without inference.
+`--check-only` to validate checkpoint compatibility without inference. The
+evaluator defaults to the selected student state; pass `--state model_ema` for
+the contemporaneous EMA state.
 
 ## Architecture and Training