Deep reinforcement learning for active breast lesion detection from DCE-MRI

Gabriel Maicas, Gustavo Carneiro, Andrew P. Bradley, Jacinto C. Nascimento, Ian Reid

2017

Deep reinforcement learning agent searching for breast lesions in DCE-MRI volumes

Publication summary

This MICCAI 2017 paper presents an automated breast lesion detection method for dynamic contrast-enhanced MRI using deep reinforcement learning. The method is designed to reduce inference time compared with exhaustive lesion search methods while retaining competitive detection accuracy.

The system trains an artificial agent using a deep Q-network to progressively focus attention on suspicious regions within a 3D breast MRI volume. The agent learns actions such as translation, scaling, and triggering the end of the search when a lesion is found.

Evaluated on 117 DCE-MRI patient volumes, the method achieves similar true-positive detection performance to state-of-the-art approaches while significantly reducing detection runtime.

“This work shows how reinforcement learning can make volumetric medical image search more efficient by teaching an agent where to look next instead of scanning every possible region.”

Editorial research summary

OrthoAI research content adaptation

Key findings

0.80

True positive detection rate achieved on the test set using the proposed DQN-based lesion detection approach.

92s

Mean inference time per patient, compared with 164 seconds for the multi-scale cascade and approximately 60 minutes for structured learning.

117

Dataset size used for evaluation, with 58 training patients and 59 testing patients.

Method

Deep Q-network search policy

The method frames lesion detection as a sequential decision-making problem. A deep Q-network receives an observation from the current bounding box and selects the next action to move, scale, or terminate the search.

The search process begins with broad bounding boxes and progressively narrows attention to likely lesion locations, reducing the need for exhaustive scanning.

Deep Q-network lesion search policy for DCE-MRI

Attention through translation, scale, and trigger actions

The agent learns translation actions along the x, y, and z axes, scale actions to resize the search box, and a trigger action to indicate lesion detection.

Rewards are based on improvement in Dice overlap with lesion annotations, allowing the agent to learn efficient navigation through high-dimensional MRI volumes.

Reinforcement learning attention mechanism for breast lesion detection

Key contributions

This paper introduces a reinforcement-learning-based approach for efficient breast lesion detection in high-dimensional DCE-MRI scans.

Applies deep Q-learning to active lesion detection in dynamic contrast-enhanced breast MRI.
Introduces an attention mechanism that progressively focuses the search region using learned actions.
Handles variable-size bounding boxes for lesions with diverse shape, size, and location.
Reduces inference time compared with exhaustive and structured-output lesion detection methods.
Maintains detection accuracy comparable to prior state-of-the-art approaches.
Demonstrates feasibility on a clinical dataset of 117 DCE-MRI patient volumes.

Similar publications

MICCAI 2018

Training Medical Image Analysis Systems like Radiologists

Gabriel Maicas, Andrew P. Bradley, Jacinto C. Nascimento, Ian Reid, Gustavo Carneiro

view

CVPR 2021 · IEEE/CVF Conference on Computer Vision and Pattern Recognition

DoDNet: Learning to Segment Multi-Organ and Tumors from Multiple Partially Labeled Datasets

Jianpeng Zhang, Yutong Xie, Yong Xia, Chunhua Shen

view

view all