0.90
Best AUC achieved by the proposed curriculum-learning meta-training approach for breast screening from DCE-MRI.
Research & Publications
MICCAI 2018
Training medical image analysis systems like radiologists
Gabriel Maicas, Andrew P. Bradley, Jacinto C. Nascimento, Ian Reid, Gustavo Carneiro
2018
Gabriel Maicas, Andrew P. Bradley, Jacinto C. Nascimento, Ian Reid, Gustavo Carneiro
2018
Publication summary
This paper proposes a training strategy for medical image analysis systems inspired by how radiologists are trained. Instead of learning from one large static dataset, the model is meta-trained on multiple related tasks with small training sets and increasing task difficulty.
The authors extend model-agnostic meta-learning by introducing teacher-student curriculum learning for task selection. This allows the system to focus on tasks where learning progress is still improving, better mimicking radiologist-style training programs.
The approach is evaluated on weakly supervised breast screening from dynamic contrast-enhanced MRI. It achieves stronger classification performance than DenseNet, multiple-instance learning, and multi-task learning baselines.
“This work shows that medical AI can benefit from training strategies that resemble clinical education: structured, progressive, and focused on rapid adaptation to new diagnostic tasks.”
Editorial research summary
OrthoAI research content adaptation
Key findings
+0.05
Improvement over the best baseline AUC of 0.85 from multiple-instance learning and multi-task learning.
5 tasks
Meta-training used five related breast MRI classification tasks to improve adaptation and downstream screening performance.
Method
Radiologist-inspired meta-training
The model is meta-trained across multiple classification tasks, each built from small training and validation sets. This encourages the model to learn how to adapt quickly, rather than simply fit one narrow diagnostic problem.
After meta-training, the learned initialization is fine-tuned for the target breast screening task: distinguishing healthy or benign cases from malignant cases.
Teacher-student curriculum task sampling
The paper compares random task sampling, all-task sampling, multi-armed bandit sampling, and teacher-student curriculum learning.
Curriculum learning selects tasks based on improvement in model performance, allowing the training process to emphasize tasks that still provide useful learning signal.
Key contributions
The paper introduces a training methodology for medical AI systems that more closely resembles the progressive task-based training process used in radiology education.
- Introduces meta-training for medical image analysis systems using multiple related diagnostic tasks.
- Improves MAML-style training by replacing random task selection with teacher-student curriculum learning.
- Uses weakly supervised DCE-MRI breast screening as the target clinical application.
- Demonstrates improved AUC compared with DenseNet, multiple-instance learning, and multi-task learning baselines.
- Shows that curriculum-based task selection is more effective than random sampling and multi-armed bandit selection.
- Supports the idea that medical AI systems can benefit from training processes modeled on human clinical education.
Similar publications
view all
MICCAI 2024
AdaCBM: An Adaptive Concept Bottleneck Model for Explainable and Accurate Diagnosis
Townim F. Chowdhury, Vu Minh Hieu Phan, Kewen Liao, Minh-Son To, Yutong Xie, Anton van den Hengel, Johan W. Verjans, Zhibin Liao
CVPR 2023 · IEEE/CVF Conference on Computer Vision and Pattern Recognition
PEFAT: Boosting Semi-Supervised Medical Image Classification via Pseudo-loss Estimation and Feature Adversarial Training
Qingjie Zeng, Yutong Xie, Zilin Lu, Yong Xia