Nearest Neighbor-Based Strategy to Optimize Multi-View Triplet Network for Classification of Small-Sample Medical...

Area of Science:

• Medical imaging
• Machine learning
• Computer-aided diagnosis

Background:

• Multi-view classification with limited data is a common challenge in medical machine learning (ML).
• Triplet networks offer a two-stage representation learning approach but face issues in feature verification for classifiers.
• Current distance-based metrics don't always guarantee optimal classification performance, necessitating exhaustive tuning.

Purpose of the Study:

• To develop a novel nearest-neighbor (NN) validation strategy for effective feature selection in multi-view ML.
• To provide a theoretically grounded approach for selecting optimal features and classifiers, avoiding repeated tuning.
• To enhance feature interpretability for medical experts, shifting focus from feature engineering to data specification.

Main Methods:

• Developed a novel nearest-neighbor (NN) validation strategy based on the triplet metric.
• Utilized a two-stage representation learning approach with triplet networks.
• Evaluated the strategy on real-world medical imaging tasks: radiation therapy error prediction and sarcoma survival prediction.

Main Results:

• The proposed NN validation strategy demonstrated superiority over common deep representation learning baselines like autoencoder (AE) and softmax.
• The strategy effectively identifies whether features or classifiers require improvement, streamlining the ML workflow.
• Achieved optimal feature selection, leading to improved classification performance in medical tasks.

Conclusions:

• The novel NN validation strategy offers a transparent and efficient method for feature selection in medical ML.
• This approach enhances feature interpretability, facilitating collaboration between ML experts and medical professionals.
• The strategy successfully addresses limitations in current representation learning techniques for limited-sample medical data.