Semisupervised learning using denoising autoencoders for brain lesion detection and segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning

Background:

• Brain lesion detection and segmentation are critical for diagnosis and treatment planning.
• Denoising autoencoders (DAEs) show promise in image analysis tasks.
• Limited labeled data often hinders the training of deep learning models for medical imaging.

Purpose of the Study:

• To explore the utility of denoising autoencoders (DAEs) for brain lesion detection, segmentation, and false-positive reduction.
• To evaluate the performance of stacked denoising autoencoders (SDAEs) with limited labeled data.
• To investigate a novelty detection approach using DAEs for lesion localization.

Main Methods:

• Stacked denoising autoencoders (SDAEs) were pretrained on unlabeled patient volumes and fine-tuned with limited labeled data.
• Transfer learning was employed, pretraining a network on high-grade glioma data and fine-tuning for low-grade glioma (LGG) segmentation.
• A single-layer DAE (novelty detector) was trained to reconstruct nonlesion patches, with reconstruction errors used for lesion localization.

Main Results:

• SDAE performance showed negligible loss even when fine-tuned with only 20 labeled patients.
• Successful LGG segmentation was achieved using transfer learning, demonstrating good generalization on unseen test data (BraTS 2013, BraTS 2015).
• The novelty detector accurately localized lesions and showed good segmentation performance on ischemic brain lesions from a different database.

Conclusions:

• DAEs, particularly SDAEs, are effective for brain lesion analysis with minimal labeled data.
• Transfer learning enhances segmentation performance for specific brain tumor types like LGG.
• The novelty detector approach provides a robust method for lesion localization and demonstrates cross-database generalizability.