Relevance maps: A weakly supervised segmentation method for 3D brain tumours in MRIs.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Deep convolutional neural networks (CNNs) are vital for medical image analysis but require large annotated datasets.
• Weakly-supervised methods offer a solution by extracting local information from global labels.
• Accurate segmentation of brain tumors like Glioma is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To develop a weakly-supervised pipeline for generating Relevance Maps from 3D medical images.
• To enable automated segmentation of regions of interest, specifically brain tumors, using only classification labels.
• To improve the accuracy and visualization resolution of medical image segmentation.

Main Methods:

• Proposed a weakly-supervised pipeline utilizing localized perturbations on pre-trained 3D CNNs to extract Relevance Maps.
• Introduced an optimal perturbation generation method leveraging 3D superpixels and U-net architecture.
• Employed a perturbation loss function to maximize prediction differences between unperturbed and perturbed inputs.

Main Results:

• Successfully applied the methodology to segment Glioma brain tumors in MRI scans.
• The proposed method demonstrated superior performance compared to existing techniques in Dice Similarity Coefficient for segmentation.
• Achieved enhanced resolution in visualizations, aiding in clearer interpretation of segmented regions.

Conclusions:

• The developed weakly-supervised pipeline effectively extracts critical regions in medical images, facilitating automated segmentation.
• This approach significantly reduces the need for extensive fine-grained annotations in medical imaging AI.
• The method shows promise for improving computer-aided diagnosis pipelines, particularly for brain tumor segmentation.