Figure 6. Annotation framework is generalizable and compatible with different strains and imaging scenarios.

(A) A representative image (max-projection of 3D stack) of head ganglion neurons in NeuroPAL strain OH15495. (B) (Left) comparison of prediction accuracy for various methods that use different information. CRF_ID framework that combines relative position features along with color information performs best (n = 9 animals, *p<0.05, **p<0.01, ***p<0.001, Bonferroni paired comparison test). (Right) the best performing method predicts cell identities with high accuracy. OpenWorm static atlas was used for all methods. Color atlas was built using experimental data with test data held out. Ensemble of color atlases that combine two different color matching methods were used for prediction. Accuracy results shown for top predicted labels. Experimental data comes from strain OH15495. (C) (Left) annotation framework can easily incorporate information from annotated data in the form of data-driven atlas, which improves prediction accuracy (***p<0.001, Bonferroni paired comparison test). Prediction was performed using leave-one-out data-driven atlases for both positional relationship features and color. Accuracy shown for top predicted labels. Ensemble of color atlases that combine two different color matching methods were used for prediction. (Right) accuracy achieved by top, top 3, and top 5 labels. Experimental data comes from strain OH15495. Top, middle, and bottom lines in box plot indicate 75th percentile, median and 25th percentile of data, respectively. (D) An example image of head ganglion neurons in NeuroPAL strain for rotated animal (nematode lying on DV axis). In contrast, animal lying on the LR axis is shown below. The locations of RMDVL/R, AVEL/R cells in the two images are highlighted for contrasts. Dashed ellipses indicate positions of cells in retrovesicular ganglion, showing that the rotated animal is not rigidly rotated. Experimental data comes from strain OH15495. (E) Top-label prediction accuracies for non-rigidly rotated animal. n = 7 animals. Experimental data comes from strain OH15495 and OH15500. Prediction was performed using leave-one-out data-driven atlases for both positional relationship features and color. Accuracy shown for top predicted labels. Ensemble of color atlases that combine two different color matching methods were used for prediction.

Figure 6—figure supplement 1. Additional results on prediction performance of CRF_ID method on NeuroPAL data: comparison against registration method and utility of ensemble of color atlases.

(A) Comparing accuracy of top 3 and top 5 identities predicted by different methods show CRF_ID framework with pairwise positional relationship features performs better than registration method (top three identities case). All methods used the same leave-one-out color atlas (see Appendix 1–Extended methods S2.3). Experimental data comes from OH15495 strain. n = 9 animals. All comparisons performed with Bonferroni paired comparison test. Top, middle, and bottom lines in box plot indicate 75th percentile, median, and 25th percentile of data respectively. (B) Prediction accuracy of CRF_ID framework on experimental datasets across different kinds of data-driven atlases. ‘All’ atlas includes positional relationships and color information from all datasets including test dataset. For ‘All color, leave-one-out pos.’ atlas, test dataset is held out from positional relationships atlas only. For ‘All pos., leave-one-out-color’ atlas, test dataset is held out from color atlas only. For ‘Both leave-one-out’ atlas, test dataset is held out from both positional relationship and color atlases. Experimental data comes from OH15495 strain. n = 9 animals. Top, middle, and bottom lines in box plot indicate 75th percentile, median, and 25th percentile of data, respectively. (C) Effect of different color distribution alignment methods on prediction accuracy. ‘Both leave-one-out’ case is the baseline case that uses leave-one-out atlases for both positional relationships and color, and color atlases are built by simple aggregation of RGB values across datasets. ‘norm’ indicates normalization of color channels, ‘histm’ indicates histogram matching of training datasets (images used to build atlas) to test dataset. ‘colconst’ indicates color invariant transformation applied to images. ‘norm + histm’ indicates normalization of color channels and then histogram matching of training images to test image. ‘colconst + histm’ indicates color invariant transformation and then histogram matching of training images to test image. ‘colconst + norm + histm’ indicates color normalization, subsequent color invariant transformation and finally histogram matching of training images to test image. ‘hsv + histm’ indicates using hsv color space instead of RGB color space and histogram matching. ‘contrast and gamma’ indicates contrast and gamma adjustment of image channels. ‘contrast and gamma + histm’ indicates contrast and gamma adjustment of image channels and subsequent histogram matching. ‘contrast and gamma + colconst + histm’ indicates contrast and gamma adjust of image channels, subsequent color invariant transformation and finally histogram matching. See Appendix 1–Extended methods S2.4 for more details on methods. Experimental data comes from OH15495 strain. n = 9 animals. Top, middle, and bottom lines in box plot indicate 75th percentile, median, and 25th percentile of data, respectively. (D) Comparison of prediction accuracy of CRF_ID framework across different kinds of data-driven atlases used for prediction. Test dataset was held out from data-driven atlases. ‘Pos. atlas only’ uses only positional relationship features atlas for prediction (these results are same as Figure 2A). ‘Pos. and Col. atlas’ uses positional relationship features and baseline color atlas (Appendix 1–Extended methods S2.4) built by simple aggregation of RGB values of cells in training data used to build atlas. ‘Pos. and Col. ensemble atlas’ uses ensemble of two color atlases for prediction along with positional relationship features atlas. In this case, color distributions in training images were aligned to test data using color invariants and histogram matching prior to building atlas (Appendix 1–Extended methods S2.4). (n = 9 animals, Bonferroni paired comparison test). Experimental data comes from OH15495 and OH15000 strains. Top, middle, and bottom lines in box plot indicate 75th percentile, median, and 25th percentile of data, respectively. (E) Similar to panel B for animals non-rigidly rotated about AP axis (n = 7 animals). Test dataset was held out from data-driven atlases. In this case, Positional relationship and color data-driven atlases were built using data from animals imaged in lateral orientation as well as rotated animals. Comparisons performed with Bonferroni paired comparison test. Experimental data comes from OH15495 and OH15000 strains. Top, middle, and bottom lines in box plot indicate 75th percentile, median, and 25th percentile of data, respectively.

Figure 6—figure supplement 2. Example annotations predicted by the CRF_ID framework for animals imaged lying on the LR axis.

Data comes from OH15495 strain.

Figure 6—figure supplement 3. Example annotations predicted by the CRF_ID framework for animals twisted about the anterior-posterior axis (note the anterior and lateral ganglions show clear left-right separation whereas retrovesicular ganglion instead of being in the middle is more toward one of the left or right sides).

Data comes from OH15495 and OH15500 strains.