BI-RADS CBM

---

Learning a Clinically-Relevant Concept Bottleneck for Lesion Detection in Breast Ultrasound

Github repository containing all relevant code for MICCAI 2024 submission

This repository is designed to provide implementations of the training and validation scripts for our breast ultrasound (BUS) concept bottleneck model (CBM) from the ACR Breast Imaging and Reporting Data System (BI-RADS) masses lexicon for ultrasound, for lesion detection, description, and cancer classification.

Architecture overview:

Results

Installation and system requirements

• Tested on Ubuntu 20.04.6 LTS
• Python version: 3.9.16
• To install dependencies, run:

python setup.py install

Demo

• Demo scripts are provided in the outermost folder.
• Model architectures are provided via the configs folder.
• A demo dataset is provided purely to validate model functionality, the dataset is not representative of the complete dataset used to train/evaluate the models in the manuscript.
• To validate code functionality, run sample code in notebook corresponding to desired functionality (e.g. for an example of how to load and test pretrained models)
  • model_eval.ipynb for sample COCO-style evaluation scripts
  • model_train.py for sample model training script