JAMIA Publication:The Potential and Pitfalls of using a LLM as a Clinical Assistant

Predicting Length-of-Stay & Risk of Mortality for ICU Patients

Pangaea Data Awarded Top Tier Co-Sell Partnership Status by Microsoft

Pangaea Named ‘Digital Solution of the Year’ by UK Government’s Department for International Trade

WHITEPAPER: How can Pharmaceutical & Healthcare Providers Collaborate in a Financially Sustainable & Scalable Manner to Improve Patient Outcomes?
Go back

BiCAL: Bi-directional Contrastive Active Learning for Clinical Report Generation

This paper, published in BioNLP 2024, proposes a novel active learning method to train models with less labelled data in the medical domain.

This paper, published in BioNLP 2024, investigates a new way to train large, powerful models for specialized tasks in computer vision (CV) and natural language processing (NLP) especially for the medical domain. Currently, training these models typically needs high volumes of labelled data, which is resource intensive. This publication proposes a new method called Bidirectional Contrastive Active Learning (BiCAL) to make this process more data efficient. BiCAL uses both images and text to find and select the most important data samples for labelling, which also helps to deal with issues such as class imbalance, where some categories have much more data than others. The study focuses on generating clinical reports from chest X-ray images and shows that BiCAL significantly improves the performance of these models, compared to previous state-of-the-art methods. This demonstrates BiCAL’s effectiveness in training models for specialized clinical tasks with less labelled data.

Request Publication

You have successfuly subscribed!