Journal: Bioinformatics
Article Title: Knowledge-enhanced protein subcellular localization prediction from 3D fluorescence microscope images
doi: 10.1093/bioinformatics/btaf331
Figure Lengend Snippet: The effects of asymmetric loss (ASL) and class confidence weights (CCW) on the class imbalance and weak annotation issues. (A) Results on the 11 categories with small sample sizes before and after using ASL. (B) Results of the prediction with and without using ASL or CCW. The red protein labels in the examples are not reflected in the shown single cells. Incorporating ASL and CCW enables the model to identify the categories of the single cells more accurately, even under weak annotation influence.
Article Snippet: The results indicated that the most critical components in classification performance were the ASL and the 2D branch, and CCW has the most significant impact on the OpenCell dataset, which aligns closely with its weak annotation issue. shows that incorporating ASL has enhanced the accuracy for categories with small sample sizes, demonstrating that ASL could perform well in the class imbalance situation.
Techniques: