2018 data science bowl  (Kaggle Inc)

Journal: Advanced Intelligent Systems

Article Title: π‐PhenoDrug: A Comprehensive Deep Learning‐Based Pipeline for Phenotypic Drug Screening in High‐Content Analysis

doi: 10.1002/aisy.202400635

Figure Lengend Snippet: Figure 2. Architecture and performance comparison of the deep learning-based single-cell segmentation model NUSeg. A) The UNetþþ backbone of NUSeg. The Xception block was used as the encoder for NUSeg, and scSE attention was employed in the encoder for upsampling. Performance com- parison of NUSeg with 12 deep learning supervised cell segmentation models based on the IoU, Dice coefficient, and Hausdorff distance on B–D) the BBBC039 dataset and E–G) the Kaggle 2018 Data Science Bowl dataset. Fivefold cross-validation was implemented on both datasets. Higher IoU and Dice coefficients indicate better performance. A lower Hausdorff distance indicates superior performance.

Article Snippet: Performance comparison of NUSeg with 12 deep learning supervised cell segmentation models based on the IoU, Dice coefficient, and Hausdorff distance on B–D) the BBBC039 dataset and E–G) the Kaggle 2018 Data Science Bowl dataset.

Techniques: Comparison, Blocking Assay, Biomarker Discovery

Journal: Advanced Intelligent Systems

Article Title: π‐PhenoDrug: A Comprehensive Deep Learning‐Based Pipeline for Phenotypic Drug Screening in High‐Content Analysis

doi: 10.1002/aisy.202400635

Figure Lengend Snippet: Figure 3. Construction of the cell morphological profile. A) Schematic of the implementation of condition erosion and marker-based watershed methods for single-cell identification. B) Representative images of the BBBC039 dataset (left), Kaggle 2018 Data Science Bowl dataset (middle), and A375 cells (right) from the raw images and segmentation mask to the NUSeg model-identified cells. Blue, DAPI; green, P16. C) Construction of cell morphological profiles and their application for drug activity analysis by supervised classification or unsupervised clustering approaches. The consistency of each channel was assessed after the identification of individual independent cells. Quality control and normalization of the cell phenotype matrix were then performed. A single-well profile was obtained by calculating the mean profile of cells within each well of the plate. Both classification and clustering analysis were used in drug activity assessment. Feature importance analysis was based on SHAP values and differential analysis (such as t-tests and one-way ANOVA). D) Feature plot of morphology features (area, form factor, perimeter), intensity features (mean intensity), and texture features (homogeneity, energy).

Article Snippet: Performance comparison of NUSeg with 12 deep learning supervised cell segmentation models based on the IoU, Dice coefficient, and Hausdorff distance on B–D) the BBBC039 dataset and E–G) the Kaggle 2018 Data Science Bowl dataset.

Techniques: Marker, Activity Assay, Control

Journal: Advanced Intelligent Systems

Article Title: π‐PhenoDrug: A Comprehensive Deep Learning‐Based Pipeline for Phenotypic Drug Screening in High‐Content Analysis

doi: 10.1002/aisy.202400635

Figure Lengend Snippet: Figure 2. Architecture and performance comparison of the deep learning-based single-cell segmentation model NUSeg. A) The UNetþþ backbone of NUSeg. The Xception block was used as the encoder for NUSeg, and scSE attention was employed in the encoder for upsampling. Performance com- parison of NUSeg with 12 deep learning supervised cell segmentation models based on the IoU, Dice coefficient, and Hausdorff distance on B–D) the BBBC039 dataset and E–G) the Kaggle 2018 Data Science Bowl dataset. Fivefold cross-validation was implemented on both datasets. Higher IoU and Dice coefficients indicate better performance. A lower Hausdorff distance indicates superior performance.

Article Snippet: We also implemented comparisons on the PanNuke and Kaggle 2018 Data Science Bowl datasets, which contain multiple types of images, to evaluate the generalizability of the model.

Techniques: Comparison, Blocking Assay, Biomarker Discovery

Journal: Advanced Intelligent Systems

Article Title: π‐PhenoDrug: A Comprehensive Deep Learning‐Based Pipeline for Phenotypic Drug Screening in High‐Content Analysis

doi: 10.1002/aisy.202400635

Figure Lengend Snippet: Figure 3. Construction of the cell morphological profile. A) Schematic of the implementation of condition erosion and marker-based watershed methods for single-cell identification. B) Representative images of the BBBC039 dataset (left), Kaggle 2018 Data Science Bowl dataset (middle), and A375 cells (right) from the raw images and segmentation mask to the NUSeg model-identified cells. Blue, DAPI; green, P16. C) Construction of cell morphological profiles and their application for drug activity analysis by supervised classification or unsupervised clustering approaches. The consistency of each channel was assessed after the identification of individual independent cells. Quality control and normalization of the cell phenotype matrix were then performed. A single-well profile was obtained by calculating the mean profile of cells within each well of the plate. Both classification and clustering analysis were used in drug activity assessment. Feature importance analysis was based on SHAP values and differential analysis (such as t-tests and one-way ANOVA). D) Feature plot of morphology features (area, form factor, perimeter), intensity features (mean intensity), and texture features (homogeneity, energy).

Article Snippet: We also implemented comparisons on the PanNuke and Kaggle 2018 Data Science Bowl datasets, which contain multiple types of images, to evaluate the generalizability of the model.

Techniques: Marker, Activity Assay, Control