Deep convolutional neural network-based algorithm for muscle biopsy diagnosis.

Area of Science:

• Neuropathology
• Artificial Intelligence in Medicine
• Digital Pathology

Background:

• Histopathologic analysis of muscle biopsies is crucial for diagnosing muscle diseases.
• Limited numbers of expert pathologists and small datasets for rare diseases hinder AI development.
• Artificial intelligence (AI) offers potential to improve diagnostic accessibility.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (CNN) algorithm for classifying muscle diseases.
• To differentiate between treatable idiopathic inflammatory myopathies and non-treatable hereditary muscle diseases.
• To assess the algorithm's accuracy in classifying subtypes of both disease categories.

Main Methods:

• Trained a CNN algorithm using 4041 microscopic images from 1400 hematoxylin-and-eosin-stained muscle pathology slides.
• Evaluated the algorithm's performance in differentiating major muscle disease categories and subtypes.
• Utilized visualization technology to validate algorithm predictions against physician diagnoses.

Main Results:

• The algorithm achieved an Area Under the Curve (AUC) of 0.996 in distinguishing idiopathic inflammatory myopathies from hereditary muscle diseases, surpassing physician performance.
• Accurate classification of four idiopathic inflammatory myopathy subtypes (average AUC 0.958) and seven hereditary muscle disease subtypes (average AUC 0.936) was demonstrated.
• Validation confirmed the algorithm's reliability and similarity to expert physician diagnoses.

Conclusions:

• The developed CNN algorithm reliably differentiates and classifies muscle diseases, including subtypes.
• The AI tool demonstrates potential for straightforward clinical application, addressing diagnostic challenges in muscle pathology.
• This approach can enhance diagnostic accuracy and accessibility, particularly for rare muscle diseases.