README.MD

# Project Report: QPQS Ratio from Chest X-rays

## Methodology
### Data Augmentation
The research employed data augmentation techniques, particularly CLAHE, to enhance the model's performance. However, after some experimets, augmentations such as CLAHE were found to degrade performance.

### Model Modification and Training
The last layers of pre-trained models from the TIMM (PyTorch Image Models) library were modified to suit the prediction task. The models were trained and evaluated through scripted procedures to minimize manual intervention. The dataset was split into training, validation, and testing sets with an 80-10-10 ratio.

### Data Imbalance
The dataset exhibited heavy variance, with a higher concentration of QPQS ratios within the range of 1 to 3. Beyond this range, the data points were sparse, leading to poorer performance for those QPQS values.

## Models Tested
The following models were tested:

- **DenseNet**  
    - `densenet`
- **Inception_v3**  
    - `inception_v3`

- **ResNet**  
    - `resnet18`
    - `resnet34`
    - `resnet50`

- **VGG**  
    - `vgg11`
    - `vgg13`

- **Vision Transformer (ViT)**  
    - Base Variants:  
        - `vit_base_8`
        - `vit_base_16`
        - `vit_base_32`
    - Small Variants:  
        - `vit_small_16`
        - `vit_small_32`
    - Large Variants:  
        - `vit_large_16`
        - `vit_large_32`
    - Tiny Variant:  
        - `vit_tiny`
    - Huge Variant:  
        - `vit_huge_14`

## Results

### Performance Analysis
- **CLAHE Impact:**  
Analysis revealed that using CLAHE reduced model performance across the board.

- **Augmentation Impact:**  
Models trained without any augmentations performed better than those trained with augmentations.

- **Model Selection:**  
Among all tested models, Vision Transformers (ViT) yielded the best performance. Further experimentation and parameter tuning with ViT models were undertaken to arrive at the best configuration.

### Model Performance on Data Variance
- **Lower QPQS Values (0-3.5):**  
The models demonstrated decent performance for these values due to the abundance of data points.

- **Higher QPQS Values:**  
Models showed poorer performance due to the sparse representation of these values in the dataset, impacting the model's ability to generalize.

## Conclusion
The study concluded that Vision Transformer models performed the best in predicting the QPQS ratio from chest X-rays. Models without CLAHE and other augmentations delivered better performance, highlighting the importance of data preprocessing choices in model training. The data imbalance significantly impacted the model’s performance on higher QPQS values, emphasizing the need for a more balanced dataset for future work.


## Best Model Details

- **Model:** `vit_large_patch16`
- **Training Data:** No Augmentation
- **Validation Data:** No Augmentation
- **Testing Data:** No Augmentation
- **CLAHE:** Disabled
- **Learning Rate:** `0.00005`
- **Average Error:** `0.54`