Medical-Reports-Generator

Overview

Medical Reports Generator repository contains the source code and text for a diploma thesis dealing with the problem of automatic generation of text reports for X-ray images.

Trained Models

All trained models are available here.
GPT-2 models are now also available on Hugging Face:

• czech-gpt2-oscar
• czech-gpt2-medical

Repository layout

The repository comprises of 4 parts:

• czech_gpt2 - contains code for fine-tuning GPT2 language model to any data it gets, for the diploma thesis it is used to fine-tune Czech GPT2 language model from English GPT2 using trasnfer learning
• medical_reports_generation - contains all source code and data related to the training and evaluation of the medical reports generation models
• medical_reports_translation - this repository contains code to automatic machine translation.
• vzor-dp - text part of the diploma thesis, as the thesis is written in English the text is contained in the vzor-dp/en folder as thesis.pdf

See the individual parts for more information.

Instalation

Download

git clone https://github.com/lchaloupsky/Medical-Reports-Generator.git

Python Requirements

⚠️ Each part of the repository may have different requirements, please take a look inside of them for more information.

GPU requirements

To run the training or testing procedures on the GPU, the GPU must support appropriate CUDA and cuDNN versions. Both GPT2 and Medical report generation parts were trained on CUDA 11 and cuDNN 8 versions.

More details are available inside correspoding parts.