Are deep learning classification results obtained on CT scans fair and interpretable?

Area of Science:

• Biomedical image processing
• Artificial intelligence in healthcare
• Deep learning applications

Background:

• Deep learning excels in image classification but faces challenges in medical applications.
• Current methods often prioritize accuracy over interpretability and proper data handling.
• Random data splitting in deep learning for medical imaging can lead to misleading results.

Purpose of the Study:

• To investigate the impact of patient-level data separation in deep learning for biomedical image classification.
• To address the limitations of random data splitting in training deep neural networks for medical diagnosis.
• To improve the real-world applicability and reliability of deep learning models in healthcare.

Main Methods:

• Training deep neural networks with strict patient-level separation of training, validation, and test datasets.
• Comparing the performance of models trained with patient-level separation versus traditional random shuffling.
• Utilizing heat map visualizations to analyze model focus and interpretability.

Main Results:

• Deep neural networks trained with patient-level separation maintain accuracy on new patient data.
• Models trained with random shuffling show poor performance when tested on unseen patients.
• Heat maps indicate enhanced focus on relevant features (nodules) with patient-level separation.

Conclusions:

• Strict patient-level data separation is crucial for developing reliable deep learning models in medical imaging.
• This approach enhances model generalization and interpretability, leading to more trustworthy automatic diagnosis systems.
• Future research should adopt patient-wise data handling to ensure the clinical utility of deep learning in healthcare.