Ventral rootlets - summary

[valosekj]

This issue summarizes work done by @LouisThomasLapointe on ventral rootlets. Thank you again for amazing work Louis-Thomas! 🚀

• Model 1 (described in #36, named 101) was trained on 14 manually annotated images (see V1.tsv). fold_0 was trained.
• Model 2 (described in #37, named 104) was trained on 20 images predicted by Model 1 and manually corrected (see V2.tsv). fold_0 was trained.
• Model 3 (described in #40, named 105) was trained on 26 images predicted by Model 2 and manually corrected (see V3.tsv). fold_0 and fold_all were trained.
• Model 4 (described in #43, named 106) was trained on 31 images predicted by Model 3 and manually corrected (see D5.tsv; NOTE: these 31 images are the same as for the dorsal model). fold_all ws trained for 1000 and 2000 epochs. The performance of Model 4 was similar to the Model 3.

All datasets, models, and predictions are saved at: ~/duke/projects/ml_spinal_rootlets/ventral_rootlets.