Lung Cancer Risk Estimation with Incomplete Data: A Joint Missing Imputation Perspective.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Multi-modal data offers complementary insights for clinical prediction.
• Missing data in clinical cohorts hinders multi-modal learning.
• Existing imputation methods struggle with heterogeneous or largely missing modalities.

Purpose of the Study:

• To develop an advanced imputation method for multi-modal missing data.
• To address challenges in imputing heterogeneous and extensively missing data modalities.
• To improve clinical prediction models using multi-modal data.

Main Methods:

• Proposed Conditional PBiGAN (C-PBiGAN), a generative adversarial model.
• Modeled the joint distribution of multi-modal data (image and non-image).
• Introduced a conditional latent space and class regularization loss for imputation.

Main Results:

• C-PBiGAN demonstrated significant improvements in lung cancer risk estimation.
• Achieved higher AUC values compared to existing methods on NLST and an in-house dataset.
• Outperformed the partial bidirectional generative adversarial net (PBiGAN) by +2.9% (NLST) and +4.3% (in-house).

Conclusions:

• C-PBiGAN is a novel generative adversarial approach for multi-modal missing data imputation.
• The method effectively handles missing data across heterogeneous modalities.
• C-PBiGAN enhances the accuracy of clinical prediction tasks like lung cancer risk estimation.