Self-Supervised Generalized Zero Shot Learning for Medical Image Classification Using Novel Interpretable Saliency...

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Medical image classification faces challenges with limited data for all disease classes, impacting generalized zero shot learning (GZSL) performance on novel data.
• Existing GZSL methods often rely on class attribute vectors, which are unavailable for medical imaging datasets.

Observation:

• A novel GZSL method is proposed, leveraging self-supervised learning (SSL) for representative vector selection and unseen class feature synthesis.
• The method incorporates enhanced GradCAM saliency maps to accurately highlight diseased regions, improving clustering by differentiating class saliency and ensuring semantic consistency.

Findings:

• The proposed GZSL approach effectively synthesizes features for unseen classes and generates accurate saliency maps without requiring class attribute vectors.
• This method demonstrates superior performance compared to state-of-the-art SSL-based GZSL techniques on both natural and medical image datasets.
• Ablation studies confirm the significant contribution of individual loss terms to the overall method efficacy.

Implications:

• This work offers a robust solution for medical image classification in data-scarce scenarios, enabling the identification of rare or novel diseases.
• The developed technique advances GZSL capabilities in specialized domains like healthcare, potentially leading to improved diagnostic tools.
• The method's independence from attribute vectors broadens its applicability across various medical imaging modalities and classification tasks.