Weakly annotated bioimages stop relying on automatic features alone: feeding domain knowledge lifts TDP-43 mislocalization precision by up to 6.8%
Lead
A model-agnostic knowledge augmentation method lets machine learning exploit domain knowledge beyond automatic feature extraction on weakly annotated bioimaging data, and combined with masked autoencoder self-supervised learning it improves precision by 2.9% over the next-best method on three in-house microscopy datasets and by an average of 6.8% across three public cytomorphology benchmarks.
Story
Classification on weakly annotated bioimages no longer depends only on automatic feature extraction, because a model can draw on domain knowledge as well. Earlier approaches relied solely on automatic feature extraction, which produced inconsistent classification performance across weakly annotated datasets. The method improved classification performance across seven machine learning architectures, so the gain is not tied to one model structure.
Domain knowledge is added to the model as augmentation rather than replacing the model, so one augmentation can be attached to different architectures. Integrated with masked autoencoder self-supervised learning it gave the strongest overall performance, improving precision by 2.9% over the next-best method on three in-house microscopy datasets and by an average of 6.8% across three public cytomorphology benchmarks. Evaluation covered three in-house microscopy datasets and three public cytomorphology benchmarks, on nucleus-level classification of TDP-43 mislocalization in motor neurons derived from human induced pluripotent stem cells.
What to watch
The method targets weakly annotated bioimaging data such as microscopy and cytomorphology images and was validated across seven machine learning architectures; its strongest overall performance came when integrated with masked autoencoder self-supervised learning. The evaluated setting is nucleus-level classification of TDP-43 mislocalization in motor neurons derived from human induced pluripotent stem cells, using three in-house microscopy datasets and three public cytomorphology benchmarks.
Accurate quantification of the TDP-43 mislocalization phenotype at scale remains challenging and may contribute to inconsistencies across studies. Biological interpretations from applying the best-performing model include dose-dependent effects of MG132 and sodium arsenite on TDP-43 mislocalization, variable pathology in motor neurons from sporadic and familial ALS donors, and a lack of protection of edaravone and niclosamide against sodium arsenite-induced TDP-43 mislocalization.