An open codebase for enhancing transparency in deep learning-based breast cancer diagnosis utilizing CBIS-DDSM data.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Challenges in breast cancer diagnosis include opaque datasets and inaccessible code, hindering research replication.
• Publicly available mammography datasets like CBIS-DDSM are crucial but require standardized processing pipelines.
• Lack of reproducible research in computer-aided diagnosis impedes advancements in breast cancer detection.

Purpose of the Study:

• To provide a comprehensive, open-source codebase for the entire machine learning pipeline in breast cancer diagnosis.
• To address the reproducibility crisis by offering a transparent methodology for model development and evaluation.
• To enhance the accuracy of malignant case detection using machine learning on mammography data.

Main Methods:

• Utilized the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) mass subset.
• Developed a pilot codebase for image preprocessing, model development, and evaluation.
• Investigated the impact of varying input image sizes on model performance.

Main Results:

• Demonstrated that increasing input image size positively correlates with improved detection accuracy for malignant cases.
• Successfully created a reproducible end-to-end pipeline from data preprocessing to model evaluation.
• The developed codebase facilitates the integration of new advancements in breast cancer diagnosis.

Conclusions:

• The provided codebase and methodology enhance reproducibility in computer-aided breast cancer diagnosis.
• Increasing input size is a key factor in improving the accuracy of malignant detection models.
• This work aims to accelerate global software development for breast cancer diagnosis through open access and reproducible research.