Feature-aware unsupervised lesion segmentation for brain tumor images using fast data density functional transform.

Area of Science:

• Medical Imaging Analysis
• Computational Physics
• Artificial Intelligence

Background:

• Accurate segmentation of medical images is crucial for diagnosis and treatment planning.
• Current deep learning methods often require large labeled datasets and significant computational resources.
• Unsupervised methods offer a promising alternative for lesion recognition and segmentation.

Purpose of the Study:

• To develop an unsupervised method for lesion morphology recognition and segmentation using medical image matrices.
• To leverage the framework of fast data density functional transform (fDDFT) and geometric deep learning.
• To improve the efficiency and accuracy of medical image analysis.

Main Methods:

• Isomorphic mapping of gray-level medical image matrices onto fDDFT energy spaces.
• Integration of geometric deep learning and graph neural network metrics.
• Utilizing gridized density functionals with global convolutional kernels for feature extraction and boundary identification.
• Employing an AutoEncoder-assisted module to reduce computational complexity.
• Validation on diverse open-access datasets.

Main Results:

• Achieved unsupervised recognition of lesion morphology and accurate segmentation.
• Demonstrated efficient global convolutional operations with reduced complexity.
• Inference time averaged 1.76 seconds per object in large 3D datasets.
• Median Dice score exceeded 0.75, meeting standard deep learning requirements.
• Synergy of fDDFT and neural networks improved training (58%) and inference (51%) times, raising Dice score to 0.9415.

Conclusions:

• The proposed fDDFT-based approach enables effective unsupervised lesion segmentation in medical images.
• The method offers significant improvements in computational efficiency and accuracy compared to conventional deep learning models.
• This technique facilitates fast computational modeling for interdisciplinary applications and clinical investigation.