Deep learning referral suggestion and tumour discrimination using explainable artificial intelligence applied to
Hyungseob Shin1, Ji Eun Park2, Yohan Jun3
1Department of Electrical and Electronic Engineering, Yonsei University, Seoul, Korea.
European Radiology
|May 7, 2023
Summary
A deep learning system accurately suggested clinical referrals for brain lesions using multiparametric MRI, performing comparably to radiologists. This AI tool aids emergency neuroradiology by providing interpretable diagnostic insights.
Area of Science:
- Neuroradiology
- Artificial Intelligence in Medicine
- Medical Imaging Analysis
Background:
- Accurate and rapid clinical referral is crucial for intra-axial mass-like lesions (IMLLs) in emergency settings.
- Multiparametric MRI, including T1-weighted, FLAIR, and DWI sequences, provides comprehensive data for IMLL diagnosis.
Purpose of the Study:
- To develop and validate an interpretable deep learning (DL) system for suggesting clinical referrals for IMLLs using multiparametric MRI.
- To assess the performance of the DL system in comparison to radiologists for referral suggestion and tumor discrimination.
Main Methods:
- A DL system was trained on 747 patients with IMLLs across 30 diseases, utilizing pre- and post-contrast T1-weighted (T1CE), FLAIR, and diffusion-weighted imaging (DWI).
- The system was validated in an independent cohort of 130 emergency patients, comparing its referral suggestions and tumor discrimination capabilities against radiologists.
- Interpretability was achieved through multiparametric visualization of high-relevance regions using layer-wise relevance propagation.
Main Results:
- The DL system achieved 72.3% accuracy in referral suggestions, comparable to radiologists (72.6%).
- For tumor vs. non-tumor differentiation, the DL system showed an AUC of 0.90 and AUPRC of 0.94, similar to human readers (AUC 0.81-0.92).
- The DL system demonstrated decision-making based on relevant image features, with high overlap in tumor regions (Dice 0.77) but not in non-tumors (Dice 0.33).
Conclusions:
- The developed DL system can effectively triage patients with IMLLs using multiparametric MRI, aiding emergency neuroradiologic diagnoses.
- The system provides interpretability via multiparametric heatmaps, enhancing trust and understanding of its diagnostic suggestions.
- This AI approach offers a valuable tool for clinical decision support in the context of acute brain lesion evaluation.
