Implementation of model explainability for a basic brain tumor detection using convolutional neural networks on MRI...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Machine learning interpretability

Background:

• Neural networks are increasingly used in medical research for tasks like image analysis.
• Explainability of these complex models is often addressed late in the development cycle.
• This delay can hinder the identification of biases and appropriate data selection.

Purpose of the Study:

• To evaluate the benefits of incorporating explainability features early in the neural network development process.
• To assess if early explainability aids in identifying model bias and selecting optimal training data.
• To determine the impact of early explainability on the clinical integration of AI predictions.

Main Methods:

• Trained a neural network to classify MRI slices for specific brain tumors (vestibular schwannoma, glioblastoma) and absence of tumor.
• Integrated features to enhance model explainability from the early stages of training.
• Compared the development process and outcomes with traditional late-stage explainability approaches.

Main Results:

• Early implementation of explainability features facilitated the detection of potential biases within the model.
• Explainability aided in the judicious selection of appropriate training datasets.
• The process highlighted how interpretability can refine model performance and reliability.

Conclusions:

• Prioritizing model explainability in the early stages of medical neural network training is crucial.
• This proactive approach can lead to more robust and reliable AI tools for clinical use.
• Early explainability supports physician trust and facilitates the integration of AI-driven insights into patient care.