Deep Learning Algorithms for Corneal Amyloid Deposition Quantitation in Familial Amyloidosis

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Familial amyloidosis, Finnish (FAF) is a rare genetic disorder characterized by amyloid deposition in various organs, including the cornea.
• Accurate quantification of corneal amyloid deposition is crucial for understanding disease progression and its impact on vision.
• Current methods for assessing amyloid burden in corneal tissue can be labor-intensive and subjective.

Purpose of the Study:

• To train and validate deep learning algorithms for quantifying relative amyloid deposition (RAD) in corneal sections from FAF patients.
• To assess the correlation between RAD and visual acuity in these patients.

Main Methods:

• Corneal specimens from 42 FAF patients were stained with Congo red and digitally scanned.
• Deep learning algorithms were trained and validated for pixel-level classification of amyloid deposits and stromal tissue.
• RAD was quantified using the trained algorithms, and its association with visual acuity (logMAR) was analyzed.

Main Results:

• The deep learning algorithms demonstrated high performance in classifying amyloid areas (sensitivity 86%, specificity 92%) and corneal stromal areas (sensitivity 74%, specificity 82%).
• There was insufficient statistical evidence (Spearman's rank correlation, p = 0.091) to establish a significant correlation between RAD and visual acuity.
• The study achieved high accuracy in automated quantification of amyloid deposits in corneal tissue.

Conclusions:

• Deep learning algorithms are effective for precise pixel-level classification of amyloid and stromal tissue in corneal sections.
• Further algorithm development is needed to correlate earlier stages of amyloid deposition with visual acuity, potentially using clinical images.
• This AI-driven approach may also be applicable to other corneal dystrophies involving amyloid accumulation.