Explainable Reverse Verification of Goodness of Classification of MRI Images by Clinical Experts.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology Informatics

Background:

• Radiological errors, stemming from factors like poor technique or misjudgment, can lead to inaccurate ground truth (GT) in Magnetic Resonance (MR) imaging.
• Faulty class labels in biomedical datasets compromise the training and performance of Computer Aided Diagnosis (CAD) systems.
• Current datasets often rely on single-radiologist labeling, increasing the risk of inherent biases and errors.

Purpose of the Study:

• To develop and validate a method for verifying the accuracy and reliability of ground truth (GT) in biomedical datasets, specifically MR imaging.
• To simulate the impact of human error in radiological interpretation on dataset labeling for binary classification tasks.
• To enhance the robustness of CAD systems by ensuring the integrity of training data.

Main Methods:

• A hypothetical approach was employed to simulate faulty radiologist perspectives by randomly swapping class labels in MR image datasets.
• Experiments were conducted on benchmark datasets (DS-75, DS-160) and a self-collected dataset (NITR-DHH) using simulated faulty iterations.
• The accuracy of the simulated faulty iterations was assessed by comparing average classification parameters against the original, correctly labeled dataset.

Main Results:

• Simulated faulty iterations demonstrated the potential impact of labeling errors on dataset integrity.
• Comparison of classification parameters between original and faulty datasets provides a quantitative measure of GT reliability.
• The study successfully generated and analyzed datasets with induced labeling errors.

Conclusions:

• The proposed approach offers a robust method for verifying the genuineness and reliability of ground truth in MR imaging datasets.
• This technique can serve as a standard validation tool for the correctness of any biomedical dataset used in machine learning.
• Ensuring data integrity is paramount for the development of trustworthy and effective Computer Aided Diagnosis (CAD) systems.