Update README.md

README.md

@@ -40,15 +40,16 @@ This model is intended for research purposes in the field of medical imaging and
   * Foundational backbone for downstream models predicting organ anomalies, segmentation, or volume-level analysis via MIL.
 
 ## Training Data
-* **Dataset(s):** The model was trained exclusively on the train split of the [CT-RATE](https://huggingface.co/datasets/ibrahimhamamci/CT-RATE) dataset, leveraging annotations for guided cropping from [ReXGroundingCT](https://huggingface.co/datasets/rajpurkarlab/ReXGroundingCT).
+* **Dataset(s):** The model was trained exclusively on the train split of the [CT-RATE](https://huggingface.co/datasets/ibrahimhamamci/CT-RATE) dataset, leveraging annotations for targeted
+*  cropping from [ReXGroundingCT](https://huggingface.co/datasets/rajpurkarlab/ReXGroundingCT).
 * **Preprocessing:** Hounsfield Units (HU) were strictly clipped between `[-997.0, 888.0]`. These values correspond to the 0.5% and 99.5% pixel intensities of the foreground voxels calculated on a subset of the CT-RATE dataset. The clipped values were mapped to a `[0, 1]` range, followed by Z-score normalization utilizing a dataset mean of `-142.39` and standard deviation of `360.97`.
 
 ## Training Procedure
-* **Training System/Framework:** LeJEPA with a guided semi-3D cropping strategy, along with an auxiliary supervised loss. Distributed Data Parallel (DDP) was utilized for scaling.
+* **Training System/Framework:** LeJEPA with a targeted semi-3D cropping strategy, along with an auxiliary supervised loss. Distributed Data Parallel (DDP) was utilized for scaling.
 * **Hardware & Scale:** The model was trained across 8x NVIDIA H200 GPUs for a total of 50,000 steps (executed in two 25,000-step runs) with a global batch size of 512.
 * **Training Strategy:** Unlike standard 2D slice-by-slice SSL training, this model was trained with spatial and anatomical awareness. Global and local crops were sampled from within a 12mm physical slab (using dynamic Z-spacing calculations) rather than a single 2D plane. 
 * **Base Model:** Initialized from scratch using the `vit_large_patch14_dinov2` architecture defined in the `timm` library. No pre-trained weights were used.
-* **Training Objective(s):** Self-supervised LeJEPA loss combined with an auxiliary multi-hot supervised loss predicting the presence of 118 [TotalSegmentator](https://github.com/wasserth/TotalSegmentator) classes within the cropped regions.
+* **Training Objective(s):** Self-supervised LeJEPA loss combined with an auxiliary multi-hot supervised loss to "guide" representations, predicting the presence of 118 [TotalSegmentator](https://github.com/wasserth/TotalSegmentator) classes within the cropped regions.
 
 ### Training Convergence
 
@@ -58,10 +59,10 @@ This model is intended for research purposes in the field of medical imaging and
 ### Data Augmentation Pipeline
 During training, a specialized, GPU-accelerated augmentation pipeline was applied to generate the multi-crop views required for the LeJEPA architecture. 
 
-**1. Spatial & Guided Cropping**
+**1. Spatial & Targeted Cropping**
 * **Global Crops:** 2 global crops were generated per volume, sized at 224x224 pixels, with a random scale between 60% and 100% of the original image dimensions.
 * **Local Crops:** 8 local crops were generated per volume, sized at 140x140 pixels, with a random scale between 30% and 60% of the original image dimensions.
-* **Anatomy & Label Guidance:** Local crops had an 80% probability of being explicitly centered on physical anatomy (probabilistically guided by TotalSegmentator or ReXGroundingCT masks) to ensure the model learns from tissue rather than background air.
+* **Anatomy & Label Guidance:** Local crops had an 80% probability of being explicitly centered on physical anatomy (probabilistically targeted by TotalSegmentator or ReXGroundingCT masks) to ensure the model learns from tissue rather than background air.
 * **Flip & Resize:** All crops were resized using nearest-neighbor interpolation to prevent the artificial averaging of Hounsfield Units, followed by a random horizontal flip applied with a 50% probability.
 
 **2. Intensity & Noise Augmentations**