Update README.md

README.md

@@ -13,8 +13,9 @@ tags:
 - medgemma
 - medsiglip
 datasets:
-- BraTS2021
-- OpenNeuro
+- BraTS2020
+- TextBraTS2021
+- MPI-Leipzig_Mind-Brain-Body
 base_model:
 - google/medgemma-1.5-4b-it
 - google/medsiglip-448
@@ -38,8 +39,8 @@ BrainGemma3D is a **multimodal vision-language model** that generates clinically
 
 ## 🎯 Key Features
 
-- **🔬 Native 3D Processing**: Inflated 2D medical vision encoder ([MedSigLIP](https://huggingface.co/google/medsiglip-base-patch16-448)) to 3D for volumetric understanding
-- **📝 Clinical Accuracy**: 95.1% F1 score on pathology entity recognition (BraTS 2021 test set)
+- **🔬 Native 3D Processing**: Inflated 2D medical vision encoder ([MedSigLIP](https://huggingface.co/google/medsiglip-448)) to 3D for volumetric understanding
+- **📝 Clinical Accuracy**: 95.1% F1 score on pathology entity recognition (on BraTS dataset)
 - **🧭 Spatial Awareness**: 68.9% laterality F1 (correct left/right hemisphere localization)
 - **🔍 Interpretable**: LIME-based 3D attribution maps show which brain regions drive predictions
 - **🚀 Efficient**: Processes full 3D volumes with 32 compressed visual tokens
@@ -52,7 +53,7 @@ BrainGemma3D is a **multimodal vision-language model** that generates clinically
 BrainGemma3D combines:
 
 1. **3D Vision Encoder**: MedSigLIP inflated to 3D via center-frame initialization (Conv2D → Conv3D)  
-   *Base model: [google/medsiglip-base-patch16-448](https://huggingface.co/google/medsiglip-base-patch16-448)*
+   *Base model: [google/medsiglip-448](https://huggingface.co/google/medsiglip-448)*
 
 2. **Token Compressor**: 2-layer Perceiver that reduces 3D patches to 32 visual tokens
 
@@ -193,8 +194,8 @@ BrainGemma3D is trained in **three progressive stages** to prevent catastrophic
 - **Epochs**: 100
 
 **Dataset**: 
-- 369 BraTS 2021 brain tumor MRI cases with radiologist-written reports
-- 99 healthy control scans with synthetic reports
+- 369 [BraTS 2020](https://www.kaggle.com/datasets/awsaf49/brats20-dataset-training-validation) brain tumor MRI cases with radiologist-written reports from [TextBraTS 2021](https://github.com/Jupitern52/TextBraTS)
+- 99 healthy control scans with synthetic reports from [MPI-Leipzig Mind-Brain-Body](https://openneuro.org/datasets/ds000221/versions/00002)
 - Stratified group-based splits (70% train / 10% val / 20% test) to prevent patient leakage
 
 ---
@@ -256,7 +257,7 @@ weights, wvol = run_interpretability(
 
 <div align="left">
     <img src="https://cdn-uploads.huggingface.co/production/uploads/662a12d70951c58269b066fb/UkQwmZRwkn-rlNlFBNVkH.png" alt="LIME Interpretability" width="80%">
-    <p><i>Figure 3: LIME attribution maps for a BraTS sample. Red regions show supervoxels that positively contribute to pathology predictions. The model correctly focuses on tumor-affected areas in the left parietal and frontal lobes.</i></p>
+    <p><i>Figure 1: LIME attribution maps for a BraTS sample. Red regions show supervoxels that positively contribute to pathology predictions. The model correctly focuses on tumor-affected areas in the left parietal and frontal lobes.</i></p>
 </div>
 
 ---
@@ -302,7 +303,7 @@ weights, wvol = run_interpretability(
 
 ## 🏥 Clinical Validation Notes
 
-BrainGemma3D achieved **95.1% pathology F1** on the BraTS 2021 test set, but this does NOT imply clinical readiness. Key considerations:
+BrainGemma3D achieved **95.1% pathology F1** on the BraTS, but this does NOT imply clinical readiness. Key considerations:
 
 1. **Dataset Homogeneity**: BraTS contains predominantly glioblastomas — performance on other tumor types (meningiomas, metastases) is unknown
 2. **Report Quality**: Ground truth reports are from a single institution — may not generalize to other radiology practices
@@ -324,8 +325,7 @@ This project was developed by:
 
 ### Built With
 - [Google MedGemma](https://huggingface.co/google/medgemma-1.5-4b-it) — Medical domain language model
-- [Google MedSigLIP](https://huggingface.co/google/medsiglip-base-patch16-448) — Medical vision encoder
-- [BraTS 2021](https://www.med.upenn.edu/cbica/brats2021/) — Brain tumor segmentation dataset
+- [Google MedSigLIP](https://huggingface.co/google/medsiglip-448) — Medical vision encoder
 - [Hugging Face Transformers](https://huggingface.co/docs/transformers) — Model framework
 
 ---
@@ -333,4 +333,4 @@ This project was developed by:
 <div align="center">
     <p><i>Built with ❤️ for the <a href="https://www.kaggle.com/competitions/med-gemma-impact-challenge/overview">MedGemma Impact Challenge</a> 🏆</i></p>
     <p><i>Advancing Medical AI with Google's Health AI Developer Foundations</i></p>
-</div>
+</div>