Automated correction of training sets for biological image segmentation by convolutional networks
Abstract
The generation of large and clean training sets of labeled examples to serve as ground truth (GT) remains a major bottleneck for the application of supervised deep learning to biological and biomedical imaging data. Typically >10⁵ vignette-sized examples are required to train a convolutional neural network (CNN) of moderate size (within 1-10 million parameters), which often cannot be obtained without accepting substantial rates of errors in the training set, leading to degradation of classification results. This issue is particularly acute in the context of dense segmentation, where every pixels in an image are to be labeled according to a small number of relevant classes. Using the U-net as our basic CNN model, we investigate the effectiveness of an iterative algorithm based on the selection of 'confident' examples, to detect segmentation errors in binary masks delineating the cellular network of an entire epithelial surface imaged by fluorescence confocal microscopy. To this end, we perform a series of perturbation experiments on synthetic images of random cellular meshworks with known GT segmentation masks, in which a known rate of errors can be systematically introduced, and the effect of errors commonly arising in biological training sets can be systematically evaluated. We demonstrate the efficacy of our algorithm for the correction of elastic deformation errors, topological defects, and random mask permutations, with error rates up to 75%. Our experiments further show the practicality of this algorithm on real biological datasets, raising the perspective of a fully automated generation of high-quality training sets for biological segmentation.
Related
- Fully automated diaphragm segmentation on routine 2D-cine and 3D thoracic MRI: multireader validation
- Organotchi: an open-source tool for the automated segmentation, morphological characterization, and longitudinal tracking of organoids
- FDUNet-LRSA: efficient frequency-adaptive segmentation for automated crop disease inspection
- Automated Segmentation and Tractography of Brainstem and Subcortical White Matter
- Automated Volumetric Segmentation of Microaneurysms on OCT Using Artificial Intelligence
BibTeX
@misc{monvel2026automated,
title = {{Automated correction of training sets for biological image segmentation by convolutional networks}},
author = {Jacques Boutet de Monvel and Raphaël Etournay},
howpublished = {HAL},
year = {2026},
month = aug,
url = {https://pasteur.hal.science/pasteur-05625026v2},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Automated correction of training sets for biological image segmentation by convolutional networks
AU - Jacques Boutet de Monvel
AU - Raphaël Etournay
PY - 2026
DA - 2026/08/27
PB - HAL
UR - https://pasteur.hal.science/pasteur-05625026v2
AB - The generation of large and clean training sets of labeled examples to serve as ground truth (GT) remains a major bottleneck for the application of supervised deep learning to biological and biomedical imaging data. Typically >10⁵ vignette-sized examples are required to train a convolutional neural network (CNN) of moderate size (within 1-10 million parameters), which often cannot be obtained without accepting substantial rates of errors in the training set, leading to degradation of classification results. This issue is particularly acute in the context of dense segmentation, where every pixels in an image are to be labeled according to a small number of relevant classes. Using the U-net as our basic CNN model, we investigate the effectiveness of an iterative algorithm based on the selection of 'confident' examples, to detect segmentation errors in binary masks delineating the cellular network of an entire epithelial surface imaged by fluorescence confocal microscopy. To this end, we perform a series of perturbation experiments on synthetic images of random cellular meshworks with known GT segmentation masks, in which a known rate of errors can be systematically introduced, and the effect of errors commonly arising in biological training sets can be systematically evaluated. We demonstrate the efficacy of our algorithm for the correction of elastic deformation errors, topological defects, and random mask permutations, with error rates up to 75%. Our experiments further show the practicality of this algorithm on real biological datasets, raising the perspective of a fully automated generation of high-quality training sets for biological segmentation.
ER -