Fig. 3: Comparison of DeepSlice vs. human performance on slide-mounted coronal mouse-brain sections.

A Human alignment performance varied significantly by level of expertise on 7 datasets, which were randomly assigned as Validation (green), used for model refinement and development, and Test (magenta), used for assessment of final model performance. **** denotes a significance level of P ≤ 0.0001, Tukey’s repeated measures multiple comparison test. B DeepSlice Model Ensemble (DS-MEns)-predicted dorsoventral (DV, left scatterplot) and mediolateral (ML, right scatterplot) cutting angles corresponded closely with cutting angles in ‘Ground Truth’ S2P (gray, n = 1400 sections from 10 experiments) and human-aligned slide-mounted (magenta, n = 191 sections from 4 experiments) sections. Data are presented as mean values±SD. C Pooled data show DeepSlice alignment accuracy with reference to human operators with varying levels of expertise; the ultimate DeepSlice iteration approximates human expert performance and consists of an ensembled model composed from the top performing 2 CNNs and integrated cutting index metadata (refer to Table 1 for details of test data). Boxes indicate 25–75th percentile range with median values indicated by horizontal lines; 10–90 percentile range is indicated by bars. Pooled data indicate median ± interquartile range D. Exemplars from the Test dataset (with overlaid alignment predictions) illustrating approximate 75th, 50th, and 25th centile performance of the MEns-AI-CI model. Schematic in B derived from S2P volume: mouse.brain-map.org. Panel D experiment details are provided in Table 1.