Computerized Prediction of Radiological Observations Based on Quantitative Feature Analysis: Initial Experience in...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Computer Vision

Background:

• Accurate interpretation of radiological images is crucial for diagnosis.
• Bridging the gap between low-level image features and high-level semantic concepts remains a challenge.

Purpose of the Study:

• To develop a computerized framework for predicting radiological observations and associating imaging features with semantic terms.
• To create a system that provides human-interpretable explanations for computer-based diagnostic decisions.

Main Methods:

• A framework was developed to predict lesion characteristics from a region of interest (ROI).
• Statistical correlation was used to link low-level imaging features with high-level semantic terms.
• A radiological observation detection algorithm was employed, incorporating state-of-the-art features and a novel feature quantifying similar-looking lesions.

Main Results:

• The framework achieved an average prediction accuracy of 83.2% for radiological observations.
• Explicit associations between semantic concepts and low-level imaging features were derived.
• The derived associations closely correlated with human expectations.

Conclusions:

• The proposed framework effectively predicts lesion characteristics and establishes interpretable links between image features and semantic concepts.
• This system can enhance annotation of radiological images, facilitate automated reasoning for diagnosis, and improve user understanding of AI-driven decisions.
• The framework shows potential for application beyond liver lesion CT images to other imaging domains.