Deep Semantic Segmentation Feature-Based Radiomics for the Classification Tasks in Medical Image Analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiomics

Background:

• Combining radiomics with deep learning for medical image classification is emerging.
• Deep learning models struggle with effective lesion representation, overfitting, and optimal feature selection, especially for small lesions or datasets.

Purpose of the Study:

• To introduce a novel Deep Semantic Segmentation Feature-based Radiomics (DSFR) framework.
• To overcome challenges in deep learning-based radiomics, including feature representation and selection.

Main Methods:

• Developed a two-module DSFR framework: deep semantic feature extraction and feature selection.
• Utilized a trained segmentation network for hierarchical semantic feature extraction.
• Employed a novel feature similarity adaptation algorithm for optimal feature selection.

Main Results:

• DSFR framework demonstrated superior performance in clinical tasks.
• Outperformed state-of-the-art approaches in pathological grading prediction for pancreatic neuroendocrine neoplasms (pNENs).
• Showcased improved prediction of thrombolytic therapy efficacy in deep venous thrombosis (DVT).

Conclusions:

• The proposed DSFR framework offers an effective solution for medical image classification.
• DSFR enhances feature extraction and selection, leading to more reliable and accurate results.
• The framework shows significant potential for clinical applications in diverse medical imaging tasks.