Would You Ask it that Way?

Linjordet, Trond; Balog, Krisztian

doi:10.1145/3477495.3531739

Cited by 2 publications

(1 citation statement)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Nonetheless, existing ML models show high performance when trained with a large training dataset. Hence there is a trade-off between application-specific ML model performance vs. training dataset size [31, 32, 33, 34, 35].…”

Section: Methods and Dataset Creationmentioning

confidence: 99%

Small Training Dataset Convolutional Neural Networks for Application Specific Super-Resolution Microscopy

Mannam

Howard

2022

Preprint

View full text Add to dashboard Cite

Machine learning (ML) models based on convolutional neural networks (CNNs) have been used to significantly increase microscopy resolution, speed (signal-to-noise ratio), and data interpretation. The bottleneck in developing effective ML systems is often the need to acquire large datasets to train the neural network. This paper demonstrates how adding a dense encoder-decoder block can be useful to effectively train a CNN that provides super-resolution images from conventional diffraction-limited microscopy images when trained using a small dataset containing 15 field-of-views (FOVs). DenseED blocks use a dense layer that concatenates features from the previous convolutional layer to the next convolutional layer. Demonstrate using DenseED blocks in fully convolutional networks (FCNs) to estimate the super-resolution images when trained with a small training dataset (15 FOVs) of human cells from the Widefield2SIM dataset and the fluorescent-labeled fixed bovine pulmonary artery endothelial cells (BPAE samples). Conventional ML models without DenseED blocks trained on small datasets fail to accurately estimate super-resolution images while models, including the DenseED blocks can. The average resolution and peak signal-to-noise ratio (PSNR) improvements achieved using DenseED blocks in FCNs when trained with 15 FOVs are 2 times and ~3.2 dB, respectively. In addition, we evaluated various configurations of target image generation methods (experimentally captured target and computationally generated target) that are used to train the FCNs with and without DenseED blocks and showed with DenseED blocks outperform compared to simple FCNs without DenseED blocks. Hence, the proposed approach indicates that microscopy applications can use DenseED blocks to train on smaller datasets that are application specific/experimental modality specific imaging platforms such as MRI/X-ray and other in vivo imaging modalities.

show abstract

Section: Methods and Dataset Creationmentioning

confidence: 99%