Discovering Digital Tumor Signatures-Using Latent Code Representations to Manipulate and Classify Liver Lesions.

Area of Science:

• Oncological Medical Imaging
• Deep Learning
• Medical Image Analysis

Background:

• Generative deep learning (DL) enables unsupervised learning of latent representations.
• These representations, termed "digital tumor signatures," can potentially aid in oncological medical imaging tasks.
• Current methods may benefit from advanced unsupervised learning for lesion characterization and data augmentation.

Purpose of the Study:

• To investigate the utility of digital tumor signatures for differentiating liver lesions from normal tissue.
• To explore the application of these signatures in generating synthetic CT images with user-defined liver tumor insertion and removal.
• To evaluate the performance of an unsupervised implicit autoencoder model for these tasks.

Main Methods:

• Utilized an implicit autoencoder, combining autoencoder and generative adversarial network (GAN)-like components.
• Trained the model on abdominal CT scans, requiring minimal data for synthetic image generation.
• Employed Principal Component Analysis (PCA) embedding for latent representation analysis and machine learning classifiers (LinearSVM) for tissue differentiation.

Main Results:

• Demonstrated successful generation of synthetic liver CT images with realistic tumor lesion insertion/removal, indistinguishable from real images by expert radiologists.
• Achieved high accuracy (up to 97%), sensitivity (up to 95%), and specificity (up to 99%) in discriminating between liver lesions and normal tissue using digital tumor signatures.
• Showcased the effectiveness of unsupervised learning with minimal data for complex medical imaging applications.

Conclusions:

• The proposed unsupervised learning paradigm effectively generates synthetic medical images and facilitates targeted manipulation of liver lesions.
• "Digital tumor signatures" derived from deep learning offer a powerful tool for differentiating cancerous lesions from healthy tissue in CT scans.
• This approach holds significant potential for advancing oncological imaging analysis and data augmentation strategies.