Tatsuhiko Naito1, Yu Nagashima2, Kenichiro Taira2, Naohiro Uchio2, Shoji Tsuji2, Jun Shimizu2. 1. The Department of Neurology, The University of Tokyo Hospital, 1138655, 7-3-1 Hongo, Bunkyo-ku, Tokyo, Japan. Electronic address: tanaitou-tky@umin.ac.jp. 2. The Department of Neurology, The University of Tokyo Hospital, 1138655, 7-3-1 Hongo, Bunkyo-ku, Tokyo, Japan.
Abstract
BACKGROUND: The morphometric analysis of myelinated nerve fibers of peripheral nerves in cross-sectional optical microscopic images is valuable. Several automated methods for nerve fiber identification and segmentation have been reported. This paper presents a new method that uses a deep learning model of a convolutional neural network (CNN). We tested it for human sural nerve biopsy images. METHODS: The method comprises four steps: normalization, clustering segmentation, myelinated nerve fiber identification, and clump splitting. A normalized sample image was separated into individual objects with clustering segmentation. Each object was applied to a CNN deep learning model that labeled myelinated nerve fibers as positive and other structures as negative. Only positives proceeded to the next step. For pretraining the model, 70,000 positive and negative data each from 39 samples were used. The accuracy of the proposed algorithm was evaluated using 10 samples that were not part of the training set. A P-value of <0.05 was considered statistically significant. RESULTS: The total true-positive rate (TPR) for the detection of myelinated fibers was 0.982, and the total false-positive rate was 0.016. The defined total area similarity (AS) and area overlap error of segmented myelin sheaths were 0.967 and 0.068, respectively. In all but one sample, there were no significant differences in estimated morphometric parameters obtained from our method and manual segmentation. COMPARISON WITH EXISTING METHODS: The TPR and AS were higher than those obtained using previous methods. CONCLUSIONS: High-performance automated identification and segmentation of myelinated nerve fibers were achieved using a deep learning model.
BACKGROUND: The morphometric analysis of myelinated nerve fibers of peripheral nerves in cross-sectional optical microscopic images is valuable. Several automated methods for nerve fiber identification and segmentation have been reported. This paper presents a new method that uses a deep learning model of a convolutional neural network (CNN). We tested it for human sural nerve biopsy images. METHODS: The method comprises four steps: normalization, clustering segmentation, myelinated nerve fiber identification, and clump splitting. A normalized sample image was separated into individual objects with clustering segmentation. Each object was applied to a CNN deep learning model that labeled myelinated nerve fibers as positive and other structures as negative. Only positives proceeded to the next step. For pretraining the model, 70,000 positive and negative data each from 39 samples were used. The accuracy of the proposed algorithm was evaluated using 10 samples that were not part of the training set. A P-value of <0.05 was considered statistically significant. RESULTS: The total true-positive rate (TPR) for the detection of myelinated fibers was 0.982, and the total false-positive rate was 0.016. The defined total area similarity (AS) and area overlap error of segmented myelin sheaths were 0.967 and 0.068, respectively. In all but one sample, there were no significant differences in estimated morphometric parameters obtained from our method and manual segmentation. COMPARISON WITH EXISTING METHODS: The TPR and AS were higher than those obtained using previous methods. CONCLUSIONS: High-performance automated identification and segmentation of myelinated nerve fibers were achieved using a deep learning model.
Authors: Yu Kang T Xu; Daryan Chitsaz; Robert A Brown; Qiao Ling Cui; Matthew A Dabarno; Jack P Antel; Timothy E Kennedy Journal: Commun Biol Date: 2019-03-26