MedicalReportGeneration Save

Medical Report Generation

A base project for Medical Report Generation.

Config

• python 2.7 / tensorflow 1.8.0
• extra package: nltk, json, PIL, numpy

DataDownload

• IU X-Ray Dataset
  • The raw data is from Open-i service of the National Library, it has many public datasets.
  • The proccessed data is on Medical-Report/NLMCXR_png_pairs.zip(提取码: vdb6), you should unzip it to dir 'data/NLMCXR_png_pairs/'.
• PreTrained InceptionV3 model
  • The raw model is from TensorflowSlim Image Classification Model Library
  • The proccessed data is on Medical-Report/pretrain_model.zip(提取码: vdb6), you shold unzip it to dir 'data/pretrain_model/'.

Train

First, get post proccess data(I have done it)

• get 'data/data_entry.json', it is the report sentences.
• get 'data/train_split.json' and 'data/test_split.json', it is the ids for train/val/test.
• get 'data/vocabulary.json', it is the vocabulary extracted from report.

Second, get TFRecord files

• get 'data/train.tfrecord' and 'data/test.tfrecord'

  $ python datasets.py
  

  e.g. if you get tfrecord files, you must annotate the code for func 'get_train_tfrecord()'

Third, go train

• you can train directly.

  $ python train.py
  

• you can see the train process

  $ cd ./data
  $ tensorboard --logdir='summary'
  
  

Demo

• You could use two chest x-ray imgs to test

  $ python demo.py --img_frontal_path='./data/experiments/CXR1900_IM-0584-1001.png' --img_lateral_path='./data/experiments/CXR1900_IM-0584-2001.png' --model_path='./data/model/my-test-2500'
  

• example

  

  $ The generate report:
       no acute cardiopulmonary abnormality
       the lungs are clear
       there is no focal consolidation
       there is no focal consolidation
       there is no pneumothorax or pneumothorax
  

Framework

Core Framework

e.g.Yuan Xue et.al-Multimodal Recurrent Model with Attention for Automated Radiology Report Generation, MICCAI 2018

Experments

Metrics Results

• CNN-RNN and CNN-RNN-Att are simple base models for image caption.
• Hier-RNN is a base model for image description generation, because we have not bounding boxes, so we use visual features from CNN directly to decode the sentence word by word.
• MRNA is a base model from the MICCAI 2018 paper^[6], we use visual features from CNN to generate first sentence, then we concat visual features and semantic features(last sentence encoded from 1d-conv layers) to generate second-final sentence word by word.
• Ours are is based on MRNA, but we improve it.

e.g. I have only release code for hier rnn and MRNA because others are easy.

Details

I split train/test dataset as 2811/300, use Adam with initial learning rate is 1e-4 with 5 epoch for decay 0.9.Then I set generate max 8 sentence with max 50 words for a sentence. The word embedding size is 512 and RNN units is 512. The more details is on config.py

IU X-Rat Datasets

The raw images are 7470, but both has frontal_view and lateral_view is 3391*2. The raw report is 3927, but sentence num >= 4 is 3631, because the report sentence num between 4 and 8 occupy 90% above, so I set max sentence num = 8. Both has image-pairs and report(sentence num >= 4) is 3111.

Result Between Normal and Abnormal Reports

When I analyse the reports from datasets, I have found Normal Reports : Abnormal Reports = 2.5 : 1, unbalanced. My best result is(not release):

e.g. Total means the total Test Dataset, Normal means the normal report(no disease) of Test Dataset, Abnormal means the abnormal report(with disease or abnormality).

Summary

Process

Now, I have summarize the process of my research of Medical Report Generation.

• First, it is easy to contact this task with Image2Text Task, so I exploit the Image Captions methods to solve this task's problems, like CNN+RNN methods.
• Second, I found that Image Captions method can solve the one sentence(short), but this task has many sentences. So I use Image Paragraph Description Generation methods, like CNN+Hierarchical RNN.
• Next, I found the reports of this task has the Impression and Findings description, so I exploit QA + Hierarchical RNN method to solve this task's problems.
• Finally, I found that language informations are more important than image infos because small scale dataset, interesting.

Problems

There are many challenges for this task, I refer to some points of ^[1].

• Very Small Medical Data, most medical datasets only with images and nearly without bounding boxes and reports, so it is very very overfit.
• Very Uncertainty Report Descriptions, because different doctors have different style description for diagnosis report.
• More-Like Dense Caption Task not Story Generation, we should ground the description sentence with relevant region.
• Unsuitable Metrics, the BLEU for machine translation and CIDEr for captioning and so on are not suitable for this task.
• Impractical, up to now, there are 4-5 papers ^[5][6][7][8]. public for this task, but to be honest, they are only for papers, they do not release code.

Little Advice

• If you want to research medical report generation, you could get more data, and you could focus on the Semantic Information not Visual Information when data is small. In VQA task, someones found that Language is more useful than Image.
• You could use more stronger Language Model(BERT, ELMo or Transformer), maybe useful.

References

• [1]医学诊断报告生成论文综述
• [2]Tensorflow Model released im2text
• [3]MS COCO Caption Evaluation Tookit
• [4]TieNet Text-Image Embedding Network for Common Thorax Disease Classification and Reporting in Chest X-rays, Xiaosong Wang et at, CVPR 2018, NIH
• [5]On the Automatic Generation of Medical Imaging Reports, Baoyu Jing et al, ACL 2018, CMU
• [6]Multimodal Recurrent Model with Attention for Automated Radiology Report Generation, Yuan Xue, MICCAI 2018, PSU
• [7]Hybrid Retrieval-Generation Reinforced Agent for Medical Image Report Generation, Christy Y. Li et al, NIPS 2018, CMU
• [8]Knowledge-Driven Encode, Retrieve, Paraphrase for Medical Image Report Generation, Christy Y. Li et al, AAAI 2019, DU
• [9]A Hierarchical Approach for Generating Descriptive Image Paragraphs, Jonathan Krause et al, CVPR 2017, Stanford
• [10]Show and Tell: A Neural Image Caption Generator, Oriol Vinyals et al, CVPR 2015, Google
• [11]Show, Attend and Tell: Neural Image Caption Generation with Visual Attention, Kelvin Xu et at, ICML 2015