Research Proposal
Anmol Goel
Deep learning in Biomedical
Using Natural Language Processing
Track 1
Training word embeddings on medical corpus + using BioBert pretrained embeddings for custom NER, sentence similarity, question answering
Track 2
Visual Question Answering on medical images.
Track 1
Embeddings
- Custom Embeddings using Fasttext
- BioBERT and BioELMO pretrained on PubMed articles
- BioSentVec and BioWordVec recent advances
Named Entity Recognition
- We can create custom dataset on text corpus using Doccano
- Utilising BioBert and our custom embeddings for training BiLSTM-CNN-CRF model.
- We can also try Attention models.
Sentence Similarity
- Utilising the word and sentence embeddings for semantic sentence similarity task
- Recent Work - BioSentVec by NCBI-NLP group
Track 2
Visual Question Answering
- Using embeddings for NLP part
- Using domain specific pretrained CNN weights from NiftyNet
- Using Bayesian Networks for prediction
Misc
- Medical expert system and knowledge graph construction
- Storing in a graph database like Neo4J
- Helpful in comorbidity and Symptom-Disease relationships
- Finding co-occurrences of medical terms in PubMed articles
- Utilising Bayesian learning for uncertainty in predictions
Proposed models and architectures
- BiLSTM-CNN-CRF for NER
- Bayesian networks for CNN
- Attention based models for sentence similarity
- Ensemble methods for higher accuracy
Datasets
- https://n2c2.dbmi.hms.harvard.edu/track1 - Sentence Similarity
- https://www.imageclef.org/2019/medical/vqa - VQA
- MedQuAD - Text QA
- https://github.com/durakkerem/Medical-Question-Answer-Datasets - Text based Question Answering
deck
By anmolg
