Using Radiomic Features to Detect Anatomical Errors and Assess Deep Learning-Based Left Ventricle Segmentation in...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiomics

Background:

• Accurate left ventricle segmentation in cardiac MRI is crucial for diagnosis and computer-aided diagnosis systems.
• Deep learning methods achieve high segmentation quality but can produce undetected anatomical inconsistencies.
• Existing evaluation metrics may fail to identify subtle segmentation errors, impacting clinical interpretation.

Purpose of the Study:

• To introduce a novel application of Radiomics for assessing the anatomical quality of deep learning-based left ventricle segmentations.
• To develop and validate machine learning classifiers using radiomic features to detect anatomical errors in segmentations.
• To provide an automated quality control method for segmentation approaches, reducing reliance on manual annotations.

Main Methods:

• Extraction of radiomic features from deep learning-generated left ventricle segmentations.
• Training machine learning classifiers on these radiomic features to identify segmentations with anatomical errors.
• Extensive cross-validation on multiple public and private datasets, analyzing performance across different error severities.

Main Results:

• Radiomic-based classifiers demonstrated high performance (Accuracy, Recall, Specificity > 95%).
• Errors were detected even in less severe cases (Recall > 80%) where Dice scores indicated good quality (> 0.8).
• High generalization ability (F1-score > 0.8) across different datasets and clinical settings was observed.

Conclusions:

• Radiomic-based classifiers reliably detect anatomical errors in segmentations, serving as an effective quality control measure.
• This method reduces the need for extensive ground truth annotations for quality assessment.
• The approach offers a valuable alternative for ensuring the clinical reliability of automated segmentation techniques.