AIVE minimizes the influence of variable AI predictions. (A–D) Flow chart depicting the technical basis of AIVE, in which (A) raw 3D data from the FIB-SEM are used to (B) train and apply a classifier model to generate probability maps depicting the location of cellular membranes (dynamic range, 0–1), while (C) intensity values from the original dataset are normalized via CLAHE (dynamic range, 0–255); (D) values from B are then multiplied by values from C (see also Fig. S1 A), which isolates the voxel values belonging to a membrane to generate raw AIVE data (application of AIVE to other classes of signal provided in Fig. S1 B). (E and F) 2D slices depicting raw AIVE data generated using each algorithm shown in Fig. 1 F (3D renders of the same data provided in Fig. S2 C), and (F) averaged ΔV values for each two-way comparison between one AIVE output and each of the other outputs (both shown as a percentage of 8-bit dynamic range; 0–255). (G) Quantification of the average ΔV per slice for each two-way comparison made between the model predictions (Fig. 1, F and G) and between AIVE outputs (E and F) generated using those same model predictions (quantitation and 3D renders of blurred AI predictions shown in Fig. S1 C and Fig. S2 B, respectively). (H) 3D rendering of membrane surfaces from the test dataset (from Fig. 1 E) viewed from the front and above, indicating the positions of the static 3D probes used to benchmark membrane distance measurements (membranes nearest to each 3D probe are indicated by red shading; see Video 1 for rotation animation depicting probe positions). (I and J) Quantification of the distance between each 3D probe (shown in H) and the nearest membranes detected by the six models shown in Fig. 1 F (raw distance measurements provided in Fig. S2 A), after segmentation of the raw predictions, blurred AI predictions, or AIVE outputs, showing (I) the range (min-max) for values reported by each probe between models and (J) the relative absolute deviation in measurements from each probe. Plot markers in I and J indicate value for individual 3D probes. The interquartile range is indicated by the box, the median is indicated by the horizontal lines, and the minimum and maximum are indicated by whiskers in G, I, and J. Scale bars; A–F, 200 nm. MLP, multilayer perceptron; DT, decision table; PART, projective adaptive resonance theory.
Figure 2.

AIVE minimizes the influence of variable AI predictions. (A–D) Flow chart depicting the technical basis of AIVE, in which (A) raw 3D data from the FIB-SEM are used to (B) train and apply a classifier model to generate probability maps depicting the location of cellular membranes (dynamic range, 0–1), while (C) intensity values from the original dataset are normalized via CLAHE (dynamic range, 0–255); (D) values from B are then multiplied by values from C (see also Fig. S1 A), which isolates the voxel values belonging to a membrane to generate raw AIVE data (application of AIVE to other classes of signal provided in Fig. S1 B). (E and F) 2D slices depicting raw AIVE data generated using each algorithm shown in Fig. 1 F (3D renders of the same data provided in Fig. S2 C), and (F) averaged ΔV values for each two-way comparison between one AIVE output and each of the other outputs (both shown as a percentage of 8-bit dynamic range; 0–255). (G) Quantification of the average ΔV per slice for each two-way comparison made between the model predictions (Fig. 1, F and G) and between AIVE outputs (E and F) generated using those same model predictions (quantitation and 3D renders of blurred AI predictions shown in Fig. S1 C and Fig. S2 B, respectively). (H) 3D rendering of membrane surfaces from the test dataset (from Fig. 1 E) viewed from the front and above, indicating the positions of the static 3D probes used to benchmark membrane distance measurements (membranes nearest to each 3D probe are indicated by red shading; see Video 1 for rotation animation depicting probe positions). (I and J) Quantification of the distance between each 3D probe (shown in H) and the nearest membranes detected by the six models shown in Fig. 1 F (raw distance measurements provided in Fig. S2 A), after segmentation of the raw predictions, blurred AI predictions, or AIVE outputs, showing (I) the range (min-max) for values reported by each probe between models and (J) the relative absolute deviation in measurements from each probe. Plot markers in I and J indicate value for individual 3D probes. The interquartile range is indicated by the box, the median is indicated by the horizontal lines, and the minimum and maximum are indicated by whiskers in G, I, and J. Scale bars; A–F, 200 nm. MLP, multilayer perceptron; DT, decision table; PART, projective adaptive resonance theory.

This Feature Is Available To Subscribers Only

or Create an Account

Close Modal
Close Modal