Update README.md

README.md

@@ -4,7 +4,7 @@ base_model:
 - Qwen/Qwen3-VL-30B-A3B-Thinking
 ---
 
-This encoder is a pure vision backbone designed for medical imaging foundation models, with a focus on radiology modalities such as CT, MRI, and X‑ray. It implements efficient 3D patch embedding, rotary position encodings, scalable Transformer blocks, multi‑scale deep feature extraction, and two self‑supervised objectives tailored for medical imagery: masked image modeling (MIM) and joint embedding predictive architecture (JEPA).
+This encoder is a pure vision backbone designed for medical imaging foundation models. It implements efficient 3D patch embedding, rotary position encodings, scalable Transformer blocks, multi‑scale deep feature extraction, and two self‑supervised objectives tailored for medical imagery: masked image modeling (MIM) and joint embedding predictive architecture (JEPA).
 
 
 ## Architecture Overview
@@ -51,9 +51,9 @@ Key components and how they map to the code:
 
 ## Radiology‑centric Design Notes
 
-- **Modalities**: CT/MRI volumes (slice stacks) and X‑ray images are supported via patch tokenization
+- **Modalities**: CT volumes (slice stacks) are supported via patch tokenization
 - **Through‑plane handling**: `temporal_patch_size` acts as slice depth for 3D patching over the Z/through‑plane axis
-- **Grayscale emphasis**: Use `in_channels=1` for CT/MRI/X‑ray to align MIM reconstruction shapes
+- **Grayscale emphasis**: Use `in_channels=1` for CT to align MIM reconstruction shapes
 - **Scalability**: Attention backends support SDPA and FlashAttention‑2 for large studies and high‑res inputs
 - **Multi‑scale features**: `deepstack_visual_indexes` provide hooks for detection/segmentation heads
 
@@ -131,26 +131,23 @@ print(encoded_patches.shape)
 
 ## Recommended Radiology Settings
 - **CT chest/abdomen**: `patch_size=16`, `temporal_patch_size=16`, `in_channels=1`
-- **MRI brain**: `patch_size=16`, `temporal_patch_size=16` (or per‑sequence 2D with `temporal_patch_size=1`)
-- **X‑ray**: `patch_size=16`, `temporal_patch_size=1`, `in_channels=1`
+<!-- - **MRI brain**: `patch_size=16`, `temporal_patch_size=16` (or per‑sequence 2D with `temporal_patch_size=1`)
+- **X‑ray**: `patch_size=16`, `temporal_patch_size=1`, `in_channels=1` -->
 
 
 ## Notes
 - FlashAttention‑2 can be enabled via the attention implementation setting in the vision config
 - Ensure volume dimensions are divisible by `patch_size` and `temporal_patch_size` (or center‑crop/pad before patchify)
-- For multi‑sequence MRI or 4‑channel inputs, set `in_channels=4` and adapt reconstruction paths accordingly
+<!-- - For multi‑sequence MRI or 4‑channel inputs, set `in_channels=4` and adapt reconstruction paths accordingly -->
 
 
 ## Training Data
-CTs for lung and abdomen from TCIA.
-TODO: Details about CTs.  
-TODO: Details about CXRs.  
-TODO: Details about MRIs.  
+Proprietary dataset of 1M CT studies.  
 
 
-## Citation
+<!-- ## Citation
 
-If you use SMB‑RAD-Encoder-v1 in your research, please cite this repository.
+If you use this model in your research, please cite this repository.
 ```
 @software{standardmodel2025smbrad,
   author = {Chen, Zekai and Adam, Irsyad and Laprade, David and Brown, Kevin and others},
@@ -160,4 +157,4 @@ If you use SMB‑RAD-Encoder-v1 in your research, please cite this repository.
   journal = {Standard Model Blog},
   url = {https://huggingface.co/standardmodelbio/SMB-RAD-Encoder-v1/edit/main/README.md}
 }
-```
+``` -->