The Right Direction Needed to Develop White-Box Deep Learning in Radiology, Pathology, and Ophthalmology: A Short...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Deep learning (DL) has surged in popularity since 2012, building on classical neural network (NN) foundations.
• Rule extraction, historically used for shallow NNs, aims to solve the
• black box
• problem.
• Highly complex deep neural networks (DNNs) from DL introduce a new interpretability challenge.

Purpose of the Study:

• To review and analyze rule extraction techniques for enhancing the transparency of DL models.
• To critically assess the limitations of current DL applications in medical fields.
• To explore methods for making DNNs more interpretable, particularly in radiology, pathology, and ophthalmology.

Main Methods:

• Review of four rule extraction approaches for DNNs in computer vision.
• Discussion of limitations of DL in medical applications from an interpretability standpoint.
• Examination of DNN-to-decision tree conversion methods and their constraints.
• Description of transparent approaches for DNNs trained via deep belief networks and convolutional neural networks.

Main Results:

• Identified and reviewed key rule extraction methods applicable to DNNs.
• Highlighted fundamental limitations and criticisms of current DL in medical fields.
• Evaluated the efficacy and drawbacks of converting DNNs to decision trees.
• Presented transparent approaches for specific DNN architectures.

Conclusions:

• Addressing the
• black box
• nature of DNNs is crucial for their reliable application in medicine.
• Rule extraction and transparent approaches offer potential solutions for improving DL interpretability.
• Further research is needed to practically implement and validate these transparency methods in clinical settings.