Scholarlink

In this work, a Deep Learning-based machine vision model was developed for the detection, segmentation and counting of Neural Progenitor Cell nuclei from fluorescence microscopy images. The cells were obtained from adult mice and cultivated in vitro, with cellular nuclei labeled using DAPI dye. Convolutional neural networks for instance segmentation, specifically the Mask R-CNN model with ResNet-50 and ResNet-101 backbones, were trained to recognize the nuclei, and their results were evaluated. Nuclei labeling was implemented semi-automatically, applying a Superpixel technique and then refining the segmentations from a manual process, also using a pre-trained model, which allowed to assemble a dataset of 66 images with 6392 labels in total. The results obtained with the Resnet-50 backbone show that there is an effectiveness of 98.6% for between the specialist count and model-predicted count, in addition to having an mAP50 of 98.0%. This approach has the potential to significantly reduce the time and effort required to analyze large image sets, which is especially useful in studies that require repetitive and detailed cellular analysis.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Instance Segmentation of Neural Progenitor Cell Nuclei in Fluorescence Microscopy Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Gabriel Pérez,
  • Claudia Cecilia Russo,
  • Maria Laura Palumbo,
  • Alejandro David Moroni

摘要

In this work, a Deep Learning-based machine vision model was developed for the detection, segmentation and counting of Neural Progenitor Cell nuclei from fluorescence microscopy images. The cells were obtained from adult mice and cultivated in vitro, with cellular nuclei labeled using DAPI dye. Convolutional neural networks for instance segmentation, specifically the Mask R-CNN model with ResNet-50 and ResNet-101 backbones, were trained to recognize the nuclei, and their results were evaluated. Nuclei labeling was implemented semi-automatically, applying a Superpixel technique and then refining the segmentations from a manual process, also using a pre-trained model, which allowed to assemble a dataset of 66 images with 6392 labels in total. The results obtained with the Resnet-50 backbone show that there is an effectiveness of 98.6% for between the specialist count and model-predicted count, in addition to having an mAP50 of 98.0%. This approach has the potential to significantly reduce the time and effort required to analyze large image sets, which is especially useful in studies that require repetitive and detailed cellular analysis.