Weakly annotated bioimages get domain knowledge: seven model families improve at once
Synopsis
Automated feature extraction from weakly annotated bioimaging data performs inconsistently across datasets because models depend on the images alone. A model-agnostic knowledge augmentation method lets machine learning models exploit domain knowledge beyond automated feature extraction, improving classification across seven architectures; combined with masked autoencoder self-supervised learning it raises precision by 2.9% over the next-best method on three in-house microscopy datasets and by 6.8% on average across three public cytomorphology benchmarks.
Interpretation
Classification of weakly annotated bioimages no longer rests on automated feature extraction alone; models can draw on domain knowledge as well. Earlier approaches depended solely on automated feature extraction, which produced inconsistent classification performance across weakly annotated datasets. The knowledge augmentation method improved classification performance across seven machine learning architectures.
Domain knowledge is injected as an augmentation signal rather than by altering one network design. Because the method is model-agnostic, it can be applied across different machine learning architectures. Integrated with masked autoencoder-based 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.
The best-performing model was applied to TDP-43 mislocalization from nucleus to cytoplasm in motor neuron microscopy images. That phenotype is linked to amyotrophic lateral sclerosis, yet accurate quantification at scale remains challenging and may contribute to inconsistencies across studies. The analysis used three independent microscopy datasets of motor neurons derived from human induced pluripotent stem cells and yielded ALS-related biological interpretations, including dose-dependent effects of the proteasomal stressor MG132 and the oxidative stressor 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 SA-induced TDP-43 mislocalization.
Perspective
This report draws on summary-level material covering the knowledge augmentation method across seven architectures, its precision comparison when combined with masked autoencoder-based self-supervised learning, and results on three in-house microscopy datasets and three public cytomorphology benchmarks. The intended readers are researchers and engineers working with weakly annotated microscopy and cytomorphology images who need stable classification across datasets.
The summary-level material does not give the concrete form of the knowledge augmentation, dataset sizes or statistical tests, so the robustness of the precision gains still needs confirmation in the full method description. The difficulty of quantifying TDP-43 mislocalization is described as possibly contributing to inconsistencies across studies, and the biological interpretations come from applying the best-performing model to in-vitro microscopy images.