Weakly supervised object detection with 2D and 3D regression neural networks.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Neuroimaging

Background:

• Automated detection of multiple lesions in large medical images is challenging without location information or examples.
• Existing weakly supervised methods struggle with precise localization and identifying small lesions.

Purpose of the Study:

• To develop a novel weakly supervised detection method for brain lesions using neural networks and attention maps.
• To enable accurate lesion localization using only global image-level labels, without requiring bounding boxes or segmentation masks.

Main Methods:

• A weakly supervised detection method employing neural networks to generate attention maps from segmentation network feature maps.
• Optimization using global image-level labels and a regression objective to count lesion occurrences.
• Evaluation on MNIST-based datasets and a 3D brain scan dataset for enlarged perivascular spaces detection.

Main Results:

• The proposed method outperformed others on MNIST-based datasets.
• Weakly supervised methods achieved performance close to human intrarater agreement for enlarged perivascular spaces detection.
• The method demonstrated superior area under the curve in two regions and the lowest false positive rate.

Conclusions:

• The developed weakly supervised detection method effectively identifies brain lesion locations using attention maps.
• This approach facilitates epidemiological and clinical studies of enlarged perivascular spaces and their link to cerebrovascular diseases.