On the Importance of Diversity When Training Deep Learning Segmentation Models with Error-Prone Pseudo-Labels

Applied Sciences, Journal Year: 2024, Volume and Issue: 14(12), P. 5156 - 5156

Published: June 13, 2024

The key to training deep learning (DL) segmentation models lies in the collection of annotated data. annotation process is, however, generally expensive human resources. Our paper leverages or traditional machine methods trained on a small set manually labeled data automatically generate pseudo-labels large datasets, which are then used train so-called data-reinforced models. relevance approach is demonstrated two applicative scenarios that distinct both terms task and pseudo-label generation procedures, enlarging scope outcomes our study. experiments reveal (i) reinforcement helps, even with error-prone pseudo-labels, (ii) convolutional neural networks have capability regularize their respect labeling errors, (iii) there an advantage increasing diversity when generating either by enriching manual through accurate singular samples, considering soft per sample prior information available about certainty.

Language: Английский

---

Resolving complex cartilage structures in developmental biology via deep learning-based automatic segmentation of X-ray computed microtomography images

Scientific Reports, Journal Year: 2022, Volume and Issue: 12(1)

Published: May 24, 2022

The complex shape of embryonic cartilage represents a true challenge for phenotyping and basic understanding skeletal development. X-ray computed microtomography (μCT) enables inspecting relevant tissues in all three dimensions; however, most 3D models are still created by manual segmentation, which is time-consuming tedious task. In this work, we utilised convolutional neural network (CNN) to automatically segment the cartilaginous system represented developing nasal capsule. main challenges task stem from large size image data (over thousand pixels each dimension) relatively small training database, including genetically modified mouse embryos, where phenotype analysed structures differs norm. We propose CNN-based segmentation model optimised that trained using unique manually annotated database. was able capsule with median accuracy 84.44% (Dice coefficient). time necessary new samples shortened approximately 8 h needed mere 130 s per sample. This will greatly accelerate throughput μCT analysis elements animal developmental diseases.

Language: Английский

---

SunSCC: Segmenting, Grouping and Classifying Sunspots From Ground‐Based Observations Using Deep Learning

Journal of Geophysical Research Space Physics, Journal Year: 2023, Volume and Issue: 128(12)

Published: Nov. 28, 2023

Abstract We propose a fully automated system to detect, aggregate, and classify sunspot groups according the McIntosh scheme using ground‐based white light (WL) observations from USET facility located at Royal Observatory of Belgium. The detection uses Convolutional Neural Network (CNN), trained segmentation maps obtained with an unsupervised method based on mathematical morphology image thresholding. Given mask, mean‐shift algorithm is used aggregate individual sunspots into groups. This accounts for area each as well prior knowledge regarding shape group. A group, defined by its bounding box location Sun, finally fed CNN multitask classifier. latter predicts three components Z , p c in classification scheme. tasks are organized hierarchically mimic dependency second ( ) third first ). resulting CNN‐based more accurate than classical methods, enhancement up 16% F1 score smallest sunspots, it robust presence clouds. clustering was able separate accuracy 80%, when compared hand‐made group catalog. classifier shows comparable performances methods continuum magnetogram images recorded instruments space mission. also show that ensemble classifiers allows differentiating reliable potentially incorrect predictions.

Language: Английский

---

Neural Network Ensemble to Detect Dicentric Chromosomes in Metaphase Images

Applied Sciences, Journal Year: 2024, Volume and Issue: 14(22), P. 10440 - 10440

Published: Nov. 13, 2024

The Dicentric Chromosome Assay (DCA) is widely used in biological dosimetry, where the number of dicentric chromosomes induced by ionizing radiation (IR) exposure quantified to estimate absorbed dose an individual has received. chromosome scoring a laborious and time-consuming process which performed manually most cytogenetic biodosimetry laboratories. Further, constitutes bottleneck when several hundreds samples need be analyzed for estimation aftermath large-scale radiological/nuclear incident(s). Recently, much interest focused on automating using Artificial Intelligence (AI) tools reduce analysis time improve accuracy detection. Our study aims detect metaphase plate images ensemble artificial neural network detectors suitable datasets that present low (in this work, only 50 images). In our approach, input image first processed operators, each producing transformed image. Then, transferred specific detector trained with training set same operator Following this, provide their predictions about detected chromosomes. Finally, all are combined consensus function. Regarding operators used, were binarized separately applying Otsu Spline techniques, while morphological opening closing filters different sizes eliminate noise, isolate components, enhance structures (chromosomes) within Consensus-based decisions typically more precise than those made networks, as method can rectify certain misclassifications, assuming results correct. indicate methodology worked satisfactorily detecting majority chromosomes, remarkable classification performance even utilized. AI-based detection will beneficial rapid triage improving thereby prediction accuracy.

Language: Английский

---