Figure 8
Masking preserves most detail. Loose masks were constructed with our own implementation of cosine filtering after segmentation in Chimera (Meng et al., 2023 ) with Segger (Pintilie, 2010 ). Small (e), medium (f) and large (g) masks were created, with density reconstructed in dark blue (b), aqua (c) and light blue (d), respectively. (b) Reconstructed density inside the small mask is colored dark blue, with density from the other regions shown in gray (darker gray for medium, lighter gray for large). (c) Reconstructed density inside the medium mask is colored blue, with density from the large region in gray. (d) Reconstructed density inside the large mask is colored light blue. For the smallest mask, the validation loss in image space (lCC) and volume (real-)space (Pearson correlation, volume r2) improve over training (a), as shown visually (h) for gradient steps 10, 30, 100, 300, 1000, 3000, 10 000 and 180 000. |