Automatic Cell Counting using Active Deep Learning and Unbiased Stereology

Alahmari, Saeed; Goldgof, Dmitry B.; Hall, Lawrence O.; Mouton, Peter R.

doi:10.1109/smc.2019.8914199

Cited by 5 publications

(12 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The performance variations in accuracy across different LeNet-5 deep learning models trained on Keras is very small, Error rate (%) however, the confusion matrix shows noticeable variations on the number of cases classified correctly per class. The performance variations in segmentation is larger since the U-Net is classifying each pixel of an image as either part of a cell or background and the counting is done from the generated masks after applying a post-processing step as described in [26] [34].…”

Section: Discussionmentioning

confidence: 99%

“…The experiment to evaluate deep learning performance variations across training trials was conducted using the U-Net [35] architecture to segment cells in microscopy images, followed by a step for counting cells based on the unbiased stereology approach [36] [34]. For classifying MNIST handwritten digits we used LeNet-5 [27].…”

Section: Methodsmentioning

confidence: 99%

“…Examples of systematic errors are false assumptions related to shape, size, and orientation of objects [29]. "Here, we use the automatic version of the unbiased fractionator method [30] [31] in which a virtual volume probe disector [32] and unbiased counting rules [33] are used to estimate the total number of NeuN immunostained neurons in sections of mouse neocortex" (Page 1709) [34]. The disector frame is shown in Fig.…”

Section: Unbiased Stereologymentioning

confidence: 99%

“…We trained each model on images of 8 mice where there were 845 training images, and we tested on the ninth mouse (test mouse ID was LU3) where the total number of test images was 121. The training set was augmented using a combination of rotation and elastic deformation as detailed in [34]. Each convolutional layer was initialized in Keras using its default convolution initialization algorithm (a modified version of Glorot uniform -sometimes called Xavier uniform) [8].…”

Section: A Segmentation-based Unbiased Stereology Cell Countingmentioning

confidence: 99%

See 3 more Smart Citations

Challenges for the Repeatability of Deep Learning Models

et al. 2020

Self Cite

View full text Add to dashboard Cite

Deep learning training typically starts with a random sampling initialization approach to set the weights of trainable layers. Therefore, different and/or uncontrolled weight initialization prevents learning the same model multiple times. Consequently, such models yield different results during testing. However, even with the exact same initialization for the weights, a lack of repeatability, replicability, and reproducibility may still be observed during deep learning for many reasons such as software versions, implementation variations, and hardware differences. In this paper, we study repeatability when training deep learning models for segmentation and classification tasks using U-Net and LeNet-5 architectures in two development environments Pytorch and Keras (with TensorFlow backend). We show that even with the available control of randomization in Keras and TensorFlow, there are uncontrolled randomizations. We also show repeatable results for the same deep learning architectures using the Pytorch deep learning library. Finally, we discuss variations in the implementation of the weight initialization algorithm across deep learning libraries as a source of uncontrolled error in deep learning results.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Unbiased Stereologymentioning

confidence: 99%

Section: A Segmentation-based Unbiased Stereology Cell Countingmentioning

confidence: 99%

See 2 more Smart Citations

Challenges for the Repeatability of Deep Learning Models

et al. 2020

Self Cite

View full text Add to dashboard Cite

show abstract

“…It has also been applied to segment fluorescently-tagged neurons in the human and rat brain 32,33 . Neuron segmentation in deep learning has been compared to stereology in previous studies; thus identifying it as a valuable and reliable method for the extraction of cell counts [34][35][36][37][38] . Yet it has not been applied to segment and quantify Nissl stained pyramidal neurons in the hippocampal subfields of the human brain-a region needed for cognition and severely affected in Alzheimer's disease.…”

mentioning

confidence: 99%

Stereology neuron counts correlate with deep learning estimates in the human hippocampal subregions

Oltmer

Rosenblum

Williams

et al. 2023

Sci Rep

View full text Add to dashboard Cite

Hippocampal subregions differ in specialization and vulnerability to cell death. Neuron death and hippocampal atrophy have been a marker for the progression of Alzheimer’s disease. Relatively few studies have examined neuronal loss in the human brain using stereology. We characterize an automated high-throughput deep learning pipeline to segment hippocampal pyramidal neurons, generate pyramidal neuron estimates within the human hippocampal subfields, and relate our results to stereology neuron counts. Based on seven cases and 168 partitions, we vet deep learning parameters to segment hippocampal pyramidal neurons from the background using the open-source CellPose algorithm, and show the automated removal of false-positive segmentations. There was no difference in Dice scores between neurons segmented by the deep learning pipeline and manual segmentations (Independent Samples t-Test: t(28) = 0.33, p = 0.742). Deep-learning neuron estimates strongly correlate with manual stereological counts per subregion (Spearman’s correlation (n = 9): r(7) = 0.97, p < 0.001), and for each partition individually (Spearman’s correlation (n = 168): r(166) = 0.90, p <0 .001). The high-throughput deep-learning pipeline provides validation to existing standards. This deep learning approach may benefit future studies in tracking baseline and resilient healthy aging to the earliest disease progression.

show abstract