LEARNING TO DETECT BRAIN LESIONS FROM NOISY ANNOTATIONS.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Supervised deep learning in medical imaging requires accurate expert annotations.
• Manual voxel-level labeling of 3D medical images is challenging and prone to errors, particularly false negatives.
• Brain lesion detection in conditions like tuberous sclerosis complex is critical for diagnosis and treatment.

Purpose of the Study:

• To develop a method for training convolutional neural networks (CNNs) that can identify and correct false negatives in medical image annotations.
• To improve the accuracy of brain lesion segmentation by enhancing the quality of training data.
• To evaluate the proposed method's performance against baseline techniques.

Main Methods:

• A novel approach was developed to train CNNs for lesion segmentation.
• The method leverages CNN predictions, prediction uncertainty, and prior knowledge of lesion characteristics to identify missed lesions (false negatives).
• Identified false negatives were incorporated into the training labels to refine the model.

Main Results:

• The proposed method demonstrated superior lesion detection and segmentation accuracy compared to a baseline CNN trained on noisy labels.
• Performance improvements were observed over several alternative techniques.
• The method effectively identified and corrected false negatives in the training dataset.

Conclusions:

• The developed method successfully improves CNN-based brain lesion segmentation by addressing label imperfections (false negatives).
• This approach enhances the reliability of deep learning models in medical imaging applications.
• The findings have significant implications for automated analysis of brain lesions in pediatric neurological disorders.