Quantitative Analysis Based on Deep Neural Networks: Mild Familial Exudative Vitreoretinopathy

In the past, we investigated an automated, end-to-end system for mild familial exudative vitreoretinopathy (FEVR) in full-term infants that provided an automated diagnosis along with quantified pathological features. Using a total of 15,370 posterior pole retinal images, we constructed data sets for full-term infants, containing three separate datasets. Based on the segmented images, the classifier performed an automatic diagnosis. Our system also extracted and quantified those features: vessels density, vessels tortuosity, maximum diameter ratio, and the distance ratio of anatomical features. When we used the trained network to diagnose the Mild FEVR, we got a result with the sensitivity of 0.9474, and the specificity to 0.9984. Significant anatomic differences between Mild FEVR and Normal were identified. In Mild FEVR, vessels density was larger ( <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">12.9666± 0.9637 ( % )</tex> vs <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">12.0053 ± 0.9246(%)</tex> I), vessels tortuosity was smaller <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">(1.9098 ± 0.2647(104 cm-3 )</tex> vs <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3.3767 ± 0.2345(104cm-3)</tex> I), the maximum of the diameter ratio was higher ( <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1.8796± 0.1789</tex> vs. <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1.5075± 0.1644</tex> ), and the DMIDD ratio ( <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3.2676± 0.2638</tex> vs <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2.8199 +0.3286$</tex> ). We designed a system that automatically diagnosed mild FEVR in full-term infants. The system also provided quantified abnormal features: larger vessels density, smaller vessels tortuosity, higher maximum diameter ratio, and DM/DD ratio. These features may facilitate the early diagnosis of mild FEVR. We designed a system that automatically diagnosed mild FEVR in full-term infants. The results shown that our system automatically diagnosed mild FEVR in full-term infants, which also provided quantified abnormal features: larger vessels density, smaller vessels tortuosity, higher maximum diameter ratio, and DM/DD ratio. These features may facilitate the early diagnosis of mild FEVR.