FIGURE SUMMARY
Title

AI-assisted phenotyping in a zebrafish hypophosphatasia model enables early and precise detection of skeletal alterations

Authors
Hark, R., Zürlein, S., Nguyen, V.T., Gust, G., Hekel, L., Liedtke, D.
Source
Full text @ Sci. Rep.

Overview of the AI-based classification pipeline. (A) Images from multiple experiments are collected. A subset of 100 images was selected for manual human classification. k-fold cross-validation is performed. (B) The dataset is augmented and normalized. The BEiT model is trained using k-fold cross-validation (C) Inference is performed on the corresponding test dataset. Images were classified into three genotype classes: wildtype (WT), heterozygous (HET) and homozygous (HOM). Attention Rollout was applied to visualize the models decision making.

Generation and investigation of transgenic alpl knockout zebrafish line. (A) Schematic drawing of the genomic region on chromosome 11 deleted in alplwue7 zebrafish larvae. The inserted alpl promoter deletion is including the alpl promoter region, together with exon 1 and parts of 5’UTR and intron 1. (B) Bone and cartilage double-stainings of 120 hpf transgenic zebrafish larvae display slightly changed craniofacial development in the neurocranium and viscerocranium. Bone (red) and cartilage (blue). Two individuals representing the same genotype, but different phenotype severities are shown for hetero- and homozygotes. wildtype = alpl+/+; heterozygous = alplwue7/+; homozygous = alplwue7/wue7.

Model performance of different tested AI methods. (A) Overall classification accuracy per genotype class for the four methods (BEiT, ResNet, ViT, Human), with the dashed line indicating random-guessing. (B) Confusion matrix illustrating the distribution of predictions of the BEiT Model for the three genotype classes. (C) Receiver operating characteristic (ROC) curves for each class and method, with corresponding area-under-the-curve (AUC) values.

Correlation of AI analyzed structures by attention rollout visualizations overlaid on the original microscopic images. (A.1 & B.1) Annotated illustrations on the left side indicated key structures of the (A) neurocranium and (B) viscerocranium. (A.2-4 & B.2-4) Right side pictures highlight the spatial distribution of attention across skeletal and cartilaginous structures in different genotype groups. Cb, ceratobranchial cartilage; Ch, ceratohylal cartilage; Cl, cleithrum; Ep, ethmoid plate; M, Meckel’s cartilage; No, notochord; Op, operculum; Ot, otoliths; Ov, otic vesicle; Pc, parachordal cartilage; Pq, palatoquadrate cartilage; Th, teeth; Tr, trabecular cartilage.

Acknowledgments
This image is the copyrighted work of the attributed author or publisher, and ZFIN has permission only to display this image to its users. Additional permissions should be obtained from the applicable author or publisher of the image. Full text @ Sci. Rep.