Update README.md

README.md

@@ -25,6 +25,15 @@ This repository contains the "Better Model" (Phase 3) for breast cancer classifi
 - **Dynamic Temporal Fusion:** Uses a Multi-Stream approach where the Kinetic stream processes a variable number of post-contrast phases.
 - **Improved Temporal Aggregation:** Transitioned to **Temporal Mean Pooling** for feature fusion, providing a more robust global temporal summary compared to complex recurrent layers (LSTM/RNN) on this specific dataset volume.
 - **Explainable AI (XAI):** Full support for **3D Grad-CAM** visualization on the T2 structural stream.
+- 
+## Mathematical & Architectural Refinement
+- **From BatchNorm to InstanceNorm:** Early versions (V3) suffered from performance collapse (AUROC ~0.55) due to the conflict between a batch size of 1 and Batch Normalization layers. Switching to **Instance Normalization** resolved this by stabilizing feature maps regardless of batch size.
+- **Robust Temporal Aggregation:** To prevent the massive overfitting observed with LSTM units (~0.0003 training loss but poor validation), we implemented **Temporal Mean Pooling**. This non-learnable layer effectively summarizes the enhancement kinetics without the parameter overhead of recurrent networks.
+
+## Advanced Pre-processing & Augmentation
+- **Automatic ROI Cropping:** Uses `CropForegroundd` to focus 100% of the spatial resolution on breast tissue, removing irrelevant thoracic signals.
+- **Modality Dropout (20%):** A custom transform that randomly masks T2 or Kinetic phases during training, preventing co-adaptation and forcing the model to learn multi-modal features.
+- **Focal Loss Strategy:** Implemented to address class imbalance, specifically targeting the "Benign" class detection failure observed in baseline models.
 
 ## Model Description
 - **Developed by:** THOUAN Simon