Next, all layers were trained using SGDR for 20 cycles of 20 epochs. First, only the last layer was trained using SGDR for one cycle of 2 epochs, using a circular learning rate scheduler with a maximum learning rate of 4e-2, a minimum learning rate of 8e-3, 0.5 epochs of increasing learning rate, and 1.5 epochs of decreasing learning rate. Training images and masks were scaled up to 1024×1024 pixels and the network was further trained. The resulting weights were then saved and used as a starting point to train the network with larger, 1024×1024 images. A circular learning rate scheduler was again used, with minimum learning rates of one-twentieth of respective maximum learning rates, 2 epochs of increasing learning rates, and 18 epochs of decreasing learning rates. Differential learning rates were applied across layers, with the first third of layers having a maximum learning rate of 1e-4, the middle third having a maximum learning rate of 1e-3, and the last third having a maximum learning rate of 1e-2. Next, all layers were trained using SGDR for 1 cycle of 20 epochs. First, only the last layer was trained using stochastic gradient descent with restarts (SGDR) for 1 cycle of 8 epochs, using a circular learning rate scheduler with a maximum learning rate of 4e-2, a minimum learning rate of 8e-3, 1 epoch of increasing learning rate, and 7 epochs of decreasing learning rate ( Huang et al., 2017 Smith, 2015). A weight decay parameter of 1e-7 was used for all training. The network was first trained with center-cropped masks and images resized to 512×512 pixels, with a batch size of 4. All of our training data consisted of Schizosaccharomyces pombe colonies, with red pigment resulting from heterochromatin-mediated silencing of the ade6 + gene.
ASSERTIONERROR CELLPROFILER THREESHOLD MANUAL
This allowed us to use simpler and, when available, pre-existing annotations for training data: for the segmentation task, we used masks generated previously using the Ilastik image-processing toolkit ( Sommer et al., 2011), while for the classification task, we relied on manual labels assigned by experienced biologists to cropped images of single colonies. As insufficient training data is a common problem hampering efforts to apply deep learning in many biological domains ( Hughes et al., 2018), we opted to use a pragmatic approach, treating the segmentation and classification steps as separate problems ( Fig. 1A). While a single-step approach may be preferable from the perspective of algorithmic efficiency and speed ( Huang et al., 2016), the training data annotations are more complex, requiring both manually assigned labels and matched bounding boxes identifying the location of each colony on a plate. These two tasks could conceivably be completed either in one step, as with a single-shot detector ( Liu et al., 2016) or RetinaNet ( Lin et al., 2017), or in two separate steps, such as with a semantic segmentation, where each pixel is assigned a label such as ‘foreground’ or ‘background’, followed by classification of cropped images. We set out to leverage these recent developments to build a computational pipeline that would enable fully automated high-throughput adenine auxotrophy-based screening and quantification.
Modern machine-learning techniques such as deep learning have made huge strides in automated image classification in recent years and are beginning to be applied to previously intractable problems in the biomedical imaging domain. Nonetheless, up to now manual scoring is a common practice in the yeast community. In addition to being time-consuming and tedious, manual colony scoring may suffer from inaccuracy and irreproducibility. However, adapting this assay to quantitative high-throughput applications has proven challenging, as it requires extensive scoring of colony phenotypes by eye. This red/white colony color assay has been widely used over the last few decades for the investigation of many biological processes, such as recombination, copy number, chromosome loss, plasmid stability, prion propagation, or epigenetic gene regulation in both budding and fission yeasts. When grown under adenine-limiting conditions, adenine auxotrophs grow but accumulate a cell-limited red pigment in their vacuoles, whereas wild-type cells grow white.
For example, yeasts with mutations in the adenine biosynthetic pathway cannot grow on media lacking adenine. Auxotrophy is the inability of an organism to synthesize a particular organic compound required for its growth.
0 kommentar(er)
