In a paper published in the journal Scientific Reports, researchers discussed the vital role of microscopy in cell quantification by addressing challenges such as time consumption and errors often attributed to human intervention or automated methods in fluorescent image analysis.
Study: Machine Learning for Cell Lineage Classification in Microscopy. Image Credit: Vladimir Staykov/Shutterstock
Integration of machine learning (ML) algorithms aimed to overcome these hurdles by automating tasks and constructing predictive models from vast datasets. Leveraging unstained images enabled morphology-based classification, enhancing cell integrity assessment.
Utilizing a convolutional neural network (CNN), the researchers achieved impressive accuracy in predicting cellular lineage, albeit with slightly lower accuracies observed for specific lineages such as neuroblastoma (SH-SY5Y), liver cell treated with Mayaro virus (HUH7_mayv), and lung epithelial cell (A549).
Background
Previous research in pathology has centered on classifying cells and tissues at a subcellular level, commonly observed through microscopes. It involves identifying features like nuclear-cytoplasmic ratios and distinct nucleoli influenced by various subcellular organelles. Doctors rely on pattern recognition in microscopic images when diagnosing diseases, aided by staining techniques or optical filtering. Since the early 2000s, ML-based strategies have emerged, automating cell classification and optimizing it based on quantitative metrics.
Cell Lineage Classification
The study used CNN to quantify cell numbers in microscopy images. This regression algorithm demonstrated favorable performance and accuracy in two of three tested strains, indicating varying quantifiability among cell types. Consequently, the team focused on developing a model capable of discerning the cell lineage present in each image through a classification algorithm. CNNs were used due to their efficiency in handling image data. These networks use convolutional layers to identify particular features, followed by subsequent layers for classification tasks, all achieved without requiring intricate modifications.
The image database comprises microscopy images acquired from projects analyzed by Harmony software within an automated microscopy high-content screening (HCS) system. Specifically, phase contrast images of various cell lines were selected, including A549, HUH7_denv, 3T3, VERO6, THP1, SH-SY5Y, A172, and HUH7_mayv. Before analysis, light contrast adjustments and background corrections were performed using Harmony software to ensure image quality and consistency.
A data augmentation technique was employed, involving changes in image orientation (0°, 90°, 180°, or 270°) and scaling (75%, 50%, and 25% of the original size) to augment the image dataset. All images were resized to 200 × 200 pixels for uniformity. Additionally, filters were applied to highlight relevant characteristics, particularly for SH-SYS5, HUH7_mayv, HUH7_denv, and A549 lineages, enhancing the ability of the mode to differentiate between similar images. The Sharpen kernel yielded optimal results, accentuating image edges and improving contrast. These pre-processing steps aimed to standardize image features and improve model performance.
Cell Lineage Evaluation
The evaluation of the accuracy in classifying various cell lineages revealed promising outcomes. When the proposed algorithm with a pre-trained model using validation images was implemented, the confusion matrix predominantly exhibited correct classifications, particularly for the VERO6 and 3T3 lineages. However, the analysts noted a few misclassifications for other strains, with fewer than six images misclassified.
Despite these errors, accuracies exceeding 86% were attained for five strains, as assessed through precision, recall, and F1-score calculations. Notably, the THP and A172 strains exhibited high F1 scores of 97% and 99%, respectively, while VERO6 demonstrated exceptional accuracy with a perfect 100% score. Conversely, the SH-SY5Y, HUH7_mayv, and HUH7_denv strains yielded less accurate results, scoring below 86%.
Furthermore, the receiver operating characteristic (ROC) curve analysis validated the performance, with values close to or equal to 1.0 observed for all lineages. This robust performance across different lineages indicates the effectiveness in accurately classifying each lineage, thus enhancing confidence in its accuracy. Additionally, a comparison of regression models among the eight strains revealed varying error levels, with the A172 strain displaying the lowest mean square error (MSE) of 493.93. On the other hand, the 3T3, VERO6, and HUH7_denv strains showed higher mean square error (MSE) values, indicating comparatively lower accuracy in regression modeling.
Overall, the study underscores the efficacy of utilizing CNNs for cell lineage classification, particularly in bright field microscopy images. Automating cell counting and classification tasks provides notable advantages over manual methods, such as decreased time requirements and minimized human error.
Despite lower accuracy observed for certain strains, overall performance remained satisfactory, with potential for further refinement through preprocessing techniques. Thus, the study demonstrates the feasibility of employing computational methods for accurate and reproducible cell classification, laying the foundation for enhanced biomedical research and diagnostics.
Conclusion
In summary, evaluating various cell lineages using the proposed algorithm yielded promising results. While the model predominantly achieved correct classifications, a few misclassifications occurred for other strains. Despite these challenges, accuracies exceeding 86% were achieved for five strains, with notably high F1 scores observed for THP and A172.
Additionally, the robust performance across different lineages, validated by ROC curve analysis, highlighted the effectiveness of the model. Despite higher errors in regression modeling for some strains, the study underscored the efficacy of CNNs for cell lineage classification. Automation offered significant advantages over manual methods despite lower accuracy for certain strains. Overall, the study demonstrated the feasibility of employing computational methods for accurate cell classification, enhancing biomedical research and diagnostics.
Boost Machine Learning Trust With HEX's Human-in-the-Loop Explainability