Update README.md

README.md

@@ -25,10 +25,10 @@ base_model:
 - **Code Demo**: https://colab.research.google.com/drive/10JyssUcyqbZ9zWPop2fHwdAH5K9LpLe1?usp=sharing
 - **Paper:** http://hdl.handle.net/2117/413967
 
-## Model Results**
+## Model Results
 
 | Organ          | Dice Score (%) |
-|---------------|--------------|
+|:---------------:|:--------------:|
 | Spleen       | 97.4         |
 | Right Kidney | 96.5         |
 | Left Kidney | 96.4         |
@@ -130,39 +130,115 @@ predictor.predict_from_files(
 
 See this [**demo**](https://colab.research.google.com/drive/10JyssUcyqbZ9zWPop2fHwdAH5K9LpLe1?usp=sharing) on how to use the model and visualize its results.
 
-## Datasets
+## Training Details
+
+### Training Data
+
+The dataset is available at: https://doi.org/10.5281/zenodo.11635577
 
 GennUNet was trained using a unified dataset consisting of three large-scale abdominal organ segmentation datasets:
 - BTCV (Beyond the Cranial Vault)
 - AMOS (Abdominal Multi-Organ Segmentation)
 - TotalSegmentator
 
-The datasets were processed to remove redundant and inconsistent samples, including intensity normalization, orientation normalization, foreground cropping, and spacing standardization to ensure consistent training input.
-
 | Dataset             | Year | 5-Fold Cross-Val | Test |
-|---------------------|------|-------|---------|
+|:---------------------:|:------:|:-------:|:---------:|
 | BTCV               | 2015 | 30    | 20      |
 | AMOS               | 2022 | 272   | 200      |
 | TotalSegmentator   | 2023 | 378   | -     |
 
-## Evaluation Results
+### Training Procedure
 
-GennUNet was evaluated using a five-fold cross-validation approach, demonstrating superior segmentation performance:
+The training code is available at: https://github.com/nicoalpis/GennUNet
 
+#### Preprocessing
 
+**Patch Extraction**
 
-GennUNet demonstrates strong generalization across datasets, outperforming transformer-based models and previous state-of-the-art segmentation models.
-
-## Training Details
-
-- **Loss Function:** Dice Loss + Cross-Entropy Loss
-- **Optimizer:** Adam + Polymonial Learning Rate scheduler
-- **Initial Learning Rate:** 0.01
-- **Batch Size:** 2
-- **Augmentation:** Rotation, scaling, Gaussian noise, contrast adjustment, mirroring
-- **Training Duration:** 1000 epochs
-
+The datasets were processed to remove redundant and inconsistent samples, including intensity normalization, orientation normalization, foreground cropping, and spacing standardization to ensure consistent training input.
 
+**Data Augmentation**
+
+| Technique (MONAI)        | Probability | Range                         |
+|:------------------------:|:-----------:|:-----------------------------------------:|
+| Rotation               |     0.20    | (-0.52, 0.52)                                         |
+| Scaling               |     0.20    | (0.7, 1.4) |
+| Gaussian Noise              |     0.10    | (0, 0.1)                    |
+| Gaussian Blur             |     0.10    | (0.5, 1.0)                                          |
+| Contrast             |     0.15    | (0.75, 1.25)                                           |
+| Mirroring          |     0.50 (per axis)    |       |
+
+### Training Hyperparameters
+
+- Loss Function: Dice Loss + Cross-Entropy Loss
+- Optimizer: Adam
+- Learning Rate: 0.01
+- Weight Decay: 0.00003
+- Scheduler: PolynomialLR
+- Batch Size: 2
+- Epochs 1000
+
+  ## Evaluation
+
+The evaluation code is available at: https://github.com/nicoalpis/GennUNet
+
+### Testing Data, Factors & Metrics
+
+#### External Evaluation Data
+
+- [FLARE 2022](https://flare22.grand-challenge.org/)
+- [KiTS19](https://kits19.grand-challenge.org/)
+
+#### Metrics
+
+Dice Similarity Coefficient = (2 * TP) / (2 * TP + FP + FN)
+
+### Results
+
+**Validation**
+
+| Dataset           | Dice Score (%) |
+|:------------------:|:---------------:|
+| BTCV             | 85.97          |
+| AMOS             | 90.32         |
+| TotalSegmentator | 94.25          |
+
+**Test**
+
+| Dataset           | Dice Score (%) |
+|:------------------:|:---------------:|
+| BTCV             | 86.17          |
+| AMOS             | 90.93         |
+| FLARE 2022 | 90.43          |
+| KiTS19 | 82.07          |
+
+**Model Performance Comparison**
+
+| Method                | BTCV  | AMOS  | TotalSeg | Arch |
+|:-----------------------:|:-------:|:-------:|:----------:|:------:|
+| nnUNet (org.)        | 83.08 | 88.64 | 93.20    | CNN  |
+| nnUNet ResEnc M      | 83.31 | 88.77 | -        | CNN  |
+| nnUNet ResEnc L      | 83.35 | 89.41 | -        | CNN  |
+| nnUNet ResEnc XL     | 83.28 | 89.68 | -        | CNN  |
+| MedNeXt L k3         | 84.70 | 89.62 | -        | CNN  |
+| MedNeXt L k5         | 85.04 | 89.73 | -        | CNN  |
+| STU-Net S            | 82.92 | 88.08 | 84.72    | CNN  |
+| STU-Net B            | 83.05 | 88.46 | 87.67    | CNN  |
+| STU-Net L            | 83.36 | 89.34 | 88.92    | CNN  |
+| Swin UNETR           | 78.89 | 83.81 | 84.18    | TF   |
+| Swin UNETRV2         | 80.85 | 86.24 | -        | TF   |
+| nnFormer             | 80.86 | 81.55 | 79.26    | TF   |
+| CoTr                 | 81.95 | 88.02 | -        | TF   |
+| No-Mamba Base        | 83.69 | 89.04 | -        | CNN  |
+| U-Mamba Bot          | 83.51 | 89.13 | -        | Mam  |
+| U-Mamba Enc          | 82.41 | 88.38 | -        | Mam  |
+| A3DS SegResNet       | 80.69 | 87.27 | -    | CNN  |
+| A3DS DiNTS           | 78.18 | 82.35 | -        | CNN  |
+| A3DS SwinUNETR       | 76.54 | 85.05 | -    | TF   |
+| Ours (GennUNet)      | **85.97** | **90.32¹** | **94.25²** | CNN  |
+
+¹ Recall that the achieved results with the AMOS dataset lack 3 classes from the original dataset.  
+² The exact number of classes to which this study's results are being compared is not specified in the sources. 
 
 ## Environmental Impact