B-mode ultrasound-based CAD by learning using privileged information with dual-level missing modality completion.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Learning Using Privileged Information (LUPI) enhances B-mode ultrasound (BUS) computer-aided diagnosis (CAD) by leveraging elasticity ultrasound (EUS) data.
• Existing LUPI methods are limited to paired data and cannot address modality imbalance, hindering performance with additional single-modal BUS images.

Purpose of the Study:

• To propose a novel multi-view LUPI algorithm with Dual-Level Modality Completion (DLMC-LUPI) for supervised transfer learning.
• To improve the performance of BUS-based CAD by effectively utilizing paired ultrasound data and additional single-modal BUS images.

Main Methods:

• DLMC-LUPI implements dual-level (image and feature) modality completion for missing EUS data.
• A Variational Autoencoder (VAE) generates features, and a VAE-based Generative Adversarial Network (GAN) synthesizes EUS images, constrained by VAE-generated features.
• Multi-view LUPI is performed using feature vectors from real or pseudo images as source domains for classifier training.

Main Results:

• The proposed VAE-based GAN enhances EUS image synthesis quality by ensuring feature similarity between real and generated data.
• Experiments on two ultrasound datasets demonstrate that DLMC-LUPI outperforms existing algorithms.
• The method effectively improves the performance of single-modal BUS-based CAD.

Conclusions:

• DLMC-LUPI successfully addresses modality imbalance in ultrasound-based CAD by completing missing EUS data at both image and feature levels.
• The developed VAE-based GAN significantly improves the quality of synthesized EUS images.
• DLMC-LUPI offers a robust solution for enhancing BUS-based CAD performance, particularly in scenarios with limited paired data.