Extracting 2D weak labels from volume labels using multiple instance learning in CT hemorrhage detection.

Area of Science:

• Medical imaging analysis
• Machine learning in radiology
• Computational neuroimaging

Background:

• Multiple instance learning (MIL) is a supervised learning method for learning from hierarchical data.
• MIL is gaining traction for weak label learning but is underutilized in 3D medical imaging.
• 3D convolutional neural networks face challenges with anisotropic voxels, and patch-based methods struggle with whole-volume labels in CT scans.

Purpose of the Study:

• To apply MIL with a deep convolutional neural network to identify large hemorrhages (> 20cm³) in clinical CT head volumes.
• To develop a learned 2D model for hemorrhage detection without requiring 2D slice annotations.
• To evaluate the data requirements for generalizing MIL in CT neuroimaging by varying training data size.

Main Methods:

• Utilized MIL, treating individual CT head image volumes as bags and slices as instances.
• Employed a deep convolutional neural network framework for hemorrhage detection.
• Varied the amount of training data to assess generalization capabilities.

Main Results:

• A minimum of 400 patient image volumes were necessary for accurate per-slice hemorrhage detection.
• The leading model achieved an average true positive rate of 98.10% and true negative rate of 99.36% over five-fold cross-validation.
• An average precision of 0.9698 was recorded for the best-performing model.

Conclusions:

• MIL is a viable approach for detecting large hemorrhages in CT head imaging, creating a 2D model from 3D data.
• Sufficient training data (at least 400 volumes) is crucial for achieving accurate and generalizable results.
• The developed models and source code are available for further research in MIL for CT neuroimaging.