Annotation-efficient, patch-based, explainable deep learning using curriculum method for breast cancer detection in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Mammography is crucial for breast cancer detection.
• Interpreting mammograms poses a logistical challenge due to increasing volumes.
• Deep learning (DL) models offer potential but often require extensive annotated datasets.

Purpose of the Study:

• To develop an efficient DL model for breast cancer detection in mammograms.
• To utilize both weak (image-level) and strong (bounding box) annotations.
• To provide explainable AI (XAI) using Grad-CAM and assess it with ground truth overlap ratio.

Main Methods:

• A patch-based DL model was developed using curriculum learning, progressively increasing patch sizes.
• The model was trained with varying levels of strong supervision (0-100%) on 1976 mammograms.
• Performance was evaluated on an internal dataset and an external dataset of 4276 mammograms.

Main Results:

• Curriculum learning models (20-100% strong labels) outperformed the baseline model (0% strong labels) in F1 scores.
• F1 scores improved from 80.55% (baseline) to 83.95% (curriculum 100) on the internal dataset.
• Similar performance trends were observed on the external dataset, with F1 scores ranging from 74.65% to 78.73%.

Conclusions:

• Training DL models with curriculum learning and a patch-based approach is effective, even with limited strongly annotated data.
• This method offers satisfactory performance and explainability (XAI), addressing DL's data requirements and "black-box" nature.
• The approach shows promise for deploying DL in large-scale mammography screening programs.