No abstract

Cell painting technique provides large amount of potential information for applications such as drug discovery, bioactivity prediction and cytotoxicity assessment. However, its utility is restricted due to the requirement of advanced, costly and specific instrumentation protocols. Therefore, creating cell painted images using simple microscopic data can provide a better alternative for these applications. This study investigates the applicability of deep network-based semantic segmentation to generate cell painted images of nuclei, endoplasmic reticulum (ER) and cytoplasm from a composite image. For this, 3456 composite images from a public dataset of Broad Bioimage Benchmark collection are considered. The corresponding ground truth images for nuclei, ER and cytoplasm are generated using Otsu’s thresholding technique and used as labeled dataset. Semantic segmentation network is applied to these data and optimized using stochastic gradient descent with momentum algorithm at a learning rate of 0.01. The segmentation performance of the trained network is evaluated using accuracy, loss, mean Boundary [Formula: see text] (BF) score, Dice Index, Jaccard Index and structural similarity index. Gradient weighted Class Activation Mapping (Grad-CAM) is employed to visualize significant image regions identified by the model. Further, a cellular index is proposed as a geometrical measure which is capable of differentiating the segmented cell organelles. The trained model yields 96.52% accuracy with a loss of 0.07 for 50 epochs. Dice Index of 0.93, 0.76 and 0.75 is achieved for nuclei, ER and cytoplasm respectively. It is observed that nuclei to cytoplasm provides comparatively higher percentage change (74.56%) in the ratiometric index than nuclei to ER and ER to cytoplasm. The achieved results demonstrate that the proposed study can predict the cell painted organelles from a composite image with good performance measures. This study could be employed for generating cell painted organelles from raw microscopy images without using specific fluorescent labeling.

No abstract

Discriminating the cell organelles from microscopic images is a challenging task due to their high similarity in image appearance. In this work, an attempt has been made to differentiate nuclei, Endoplasmic Reticulum (ER) and cytoplasm using a texture pattern descriptor and Random Forest classifier. For this, Cell Painted public dataset from Broad Bioimage Benchmark collection are considered. Texture features are extracted from each image using Non Local Binary Pattern (NLBP) that captures the relationship between global pixels and sampling instances in a local neighborhood. Non local central pixels called anchors are derived from central pixels of image patches and compared with sampling instances. Binary string generated from this is encoded into 29 patterns. Statistical one-way analysis of variance (ANOVA) is performed to select significant features and are validated using Random Forest classifier. The dependency of classifier performance on the local patch radius (R) and the number of anchors (K) are also evaluated. The results indicate that 8 patterns out of 29 are showing strong inter class variability with high F value. Classification accuracy of 84% is achieved with R=3 and K=5. Experimental results demonstrate that the proposed work captures complex patterns in cell structure useful for differentiating cell components which can be employed for evaluating the cytotoxic effects in cell lines.

Differentiation of cell organelle characteristics from microscopic images is a challenging task due to its intricate structural details. In this work, an attempt has been made to categorize Endoplasmic Reticulum (ER) and cytoplasm using orthogonal Zernike moments and Multilayer Perceptron (MLP). For this, Cell painted public source dataset comprising of ER and cytoplasm are considered. Zernike moments for different orders and repetition of the azimuthal angle are extracted to characterize the shape features. The extracted features are validated using MLP classifier for differentiating ER and cytoplasm. The prediction accuracy for variations in the number of hidden layers are evaluated. The experimental results show that the accuracy varies as the size of hidden layer increases. The extracted features with MLP achieved an accuracy of 85% with a hidden layer size of 5. The receiver operating characteristic curve (ROC) demonstrates the distinguishing power of MLP classifier with AUC=0.92. This study suggests that the proposed framework can be employed for analyzing the morphological variations of cell organelles due to chemical perturbations, genome variations and cytotoxic effects using the combination of Zernike shape descriptor and MLP. The orthogonality property of Zernike shape descriptor provides independent unique features which reduce redundancy and improve prediction accuracy for large datasets.