Systematically designing better instance counting models on cell images with Neural Arithmetic Logic Units
This addresses the issue of poor generalization in automated cell counting for biomedical applications, though it is incremental as it builds on existing architectures.
The paper tackled the problem of neural networks failing to generalize to higher cell counts outside their training range by proposing the use of Neural Arithmetic Logic Units (NALU) integrated into architectures like Fully Convolutional Regression Network and U-Net, achieving better accuracy in cell counting tasks.
The big problem for neural network models which are trained to count instances is that whenever test range goes high training range generalization error increases i.e. they are not good generalizers outside training range. Consider the case of automating cell counting process where more dense images with higher cell counts are commonly encountered as compared to images used in training data. By making better predictions for higher ranges of cell count we are aiming to create better generalization systems for cell counting. With architecture proposal of neural arithmetic logic units (NALU) for arithmetic operations, task of counting has become feasible for higher numeric ranges which were not included in training data with better accuracy. As a part of our study we used these units and different other activation functions for learning cell counting task with two different architectures namely Fully Convolutional Regression Network and U-Net. These numerically biased units are added in the form of residual concatenated layers to original architectures and a comparative experimental study is done with these newly proposed changes. This comparative study is described in terms of optimizing regression loss problem from these models trained with extensive data augmentation techniques. We were able to achieve better results in our experiments of cell counting tasks with introduction of these numerically biased units to already existing architectures in the form of residual layer concatenation connections. Our results confirm that above stated numerically biased units does help models to learn numeric quantities for better generalization results.