README.md

---
datasets:
- StanfordAIMI/interpret-cxr-test-public
- StanfordAIMI/interpret-cxr-test-hidden
library_name: transformers
license: apache-2.0
---

# CXRMate-RRG24: Entropy-Augmented Self-Critical Sequence Training for Radiology Report Generation

This is an evolution of https://huggingface.co/aehrc/cxrmate developed for the [Radiology Report Generation](https://stanford-aimi.github.io/RRG24/) task of [BioNLP @ ACL 2024](https://aclweb.org/aclwiki/BioNLP_Workshop).

The leaderboard for the task can be found [here](https://vilmedic.app/misc/bionlp24/leaderboard).

For this, we proposed EAST: Entropy-Augmented Self-critical sequence Training (EAST):
 - EAST modifies [Self-Critical Sequence Training (SCST)](https://openaccess.thecvf.com/content_cvpr_2017/papers/Rennie_Self-Critical_Sequence_Training_CVPR_2017_paper.pdf) by adding entropy regularisation. 
 - Helps maintain a higher entropy in the token distribution.
 - Preventing overfitting to common phrases and ensuring a broader exploration of the vocabulary during training.
 - This was essential to handle the diversity of the radiology reports in the RRG24 datasets. 

EAST was applied to a multimodal language model with RadGraph as the reward. Other features include:
 - Token type embeddings to differentiate between findings and impression section tokens, as well as image embeddings. 
 - Special tokens (`[NF]` and `[NI]`) to handle missing *findings* and *impression* sections. 
 - Non-causal attention masking for the image embeddings and a causal attention masking for the report token embeddings.

## Paper:
https://aclanthology.org/2024.bionlp-1.8/

## Example:

```python
import torch  
from torchvision.transforms import v2
import transformers


tokenizer = transformers.AutoTokenizer.from_pretrained('aehrc/cxrmate-rrg24')
model = transformers.AutoModel.from_pretrained('aehrc/cxrmate-rrg24', trust_remote_code=True)
transforms = v2.Compose(
    [
        v2.PILToTensor(),
        v2.Grayscale(num_output_channels=3),
        v2.Resize(size=model.config.encoder.image_size, antialias=True),
        v2.CenterCrop(size=[model.config.encoder.image_size]*2),
        v2.ToDtype(torch.float32, scale=True),
        v2.Normalize(mean=model.config.encoder.image_mean, std=model.config.encoder.image_std),
    ]
)

dataset = datasets.load_dataset('StanfordAIMI/interpret-cxr-test-public')['test']

def transform_batch(batch):
    batch['images'] = [torch.stack([transforms(j) for j in i]) for i in batch['images']]
    batch['images'] = torch.nn.utils.rnn.pad_sequence(batch['images'], batch_first=True, padding_value=0.0)  
    return batch

dataset = dataset.with_transform(transform_batch)
dataloader = DataLoader(dataset, batch_size=mbatch_size, shuffle=True)
batch = next(iter(dataloader))

output_ids = model.generate(
    pixel_values=batch['images'],
    max_length=512,
    num_beams=4,
    bad_words_ids=[[tokenizer.convert_tokens_to_ids('[NF]')], [tokenizer.convert_tokens_to_ids('[NI]')]],
)
findings, impression = model.split_and_decode_sections(output_ids, tokenizer)
```

## Generate findings only:

```python
output_ids = model.generate(
    pixel_values=batch['images'],
    max_length=512,
    num_beams=4,
    bad_words_ids=[[tokenizer.convert_tokens_to_ids('[NF]')]],
    eos_token_id=tokenizer.sep_token_id,
)
findings, _ = model.split_and_decode_sections(output_ids, tokenizer)
```

## Generate impression only:

```python
output_ids = model.generate(
    pixel_values=batch['images'],
    max_length=512,
    num_beams=4,
    bad_words_ids=[[tokenizer.convert_tokens_to_ids('[NI]')]],
    input_ids=torch.tensor([[tokenizer.bos_token_id, tokenizer.convert_tokens_to_ids('[NF]'), tokenizer.sep_token_id]]*mbatch_size, device=device, dtype=torch.long),
)
_, impression = model.split_and_decode_sections(output_ids, tokenizer)
```

## Notebook example:
https://huggingface.co/aehrc/cxrmate-rrg24/blob/main/demo.ipynb

## Known issues:
There is no penalty in the reward for sampled reports that differ in length to the radiologist report. Hence, the model has learned to generate longer reports, often with repetitions. This was fixed in our recent work: https://arxiv.org/abs/2406.13181. 

## Citation:

```
@inproceedings{nicolson-etal-2024-e,
    title = "e-Health {CSIRO} at {RRG}24: Entropy-Augmented Self-Critical Sequence Training for Radiology Report Generation",
    author = "Nicolson, Aaron  and
      Liu, Jinghui  and
      Dowling, Jason  and
      Nguyen, Anthony  and
      Koopman, Bevan",
    editor = "Demner-Fushman, Dina  and
      Ananiadou, Sophia  and
      Miwa, Makoto  and
      Roberts, Kirk  and
      Tsujii, Junichi",
    booktitle = "Proceedings of the 23rd Workshop on Biomedical Natural Language Processing",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.bionlp-1.8",
    pages = "99--104",
    abstract = "The core novelty of our approach lies in the addition of entropy regularisation to self-critical sequence training. This helps maintain a higher entropy in the token distribution, preventing overfitting to common phrases and ensuring a broader exploration of the vocabulary during training, which is essential for handling the diversity of the radiology reports in the RRG24 datasets. We apply this to a multimodal language model with RadGraph as the reward. Additionally, our model incorporates several other aspects. We use token type embeddings to differentiate between findings and impression section tokens, as well as image embeddings. To handle missing sections, we employ special tokens. We also utilise an attention mask with non-causal masking for the image embeddings and a causal mask for the report token embeddings.",
}
```