Research On Automatic Recognition Of Micronucleus With Global And Local Attention Mechanism

Micronuclei(also known as satellite nuclei)are genetic material particles that are trapped outside the cell nucleus due to disorder in the chromosome separation mechanism.Since the production of micronuclei is usually associated with drugs or other toxic substances,it has long been considered as a biomarker for genetic toxicity,tumor risk,and tumor malignancy.The micronucleus test is currently one of the most popular and widely used methods for detecting micronuclei,but it is very time-consuming and prone to misdiagnosis.Therefore,research on computer-aided diagnostic systems is of great significance for micronucleus detection.Traditional methods are the most widely used in clinical diagnosis,but their performance is highly dependent on the design of manual features,and usually requires tedious steps to extract features.Deep learning methods represented by convolutional neural networks have end-to-end properties and more powerful feature representation capabilities,but their black box properties have to some extent hindered their development in medical image analysis,and improving the interpretability of deep models has become an important issue for many studies.In addition to the influence of the model itself on classification performance,the lack of available datasets due to the high sensitivity of the medical field and the rarity of diseases has also prevented many works from proceeding normally.Based on this,we start with the construction of the dataset and conducts research on image classification tasks under the background of micronucleated cells,mainly including the following three parts:(1)Construct a micronucleus cell image dataset.In response to the problem of insufficient micronucleated cell data and highly imbalanced positive and negative classes,data augmentation methods are used to properly expand the images to avoid the problem of the network failing to learn the boundaries of the two classes correctly.To explore a better dataset division method,we construct two datasets of cell images based on whether the validation set is expanded in quantity and conducts experiments on them respectively.The experimental results show that data augmentation methods that do not produce deformations are still helpful for network performance improvement on the validation set under the condition that the testing set consists of real data.(2)Propose a micronucleus cell image classification model that combines local attention mechanisms and AlexNet.Using transfer learning methods,the AlexNet pretrained on ImageNet with its fully connected layers removed is used as the backbone network.Two attention modules are used in the middle layer of the network to extract cell image features and generate attention maps that highlight regions of interest(ROIs),achieving the effect of improving the interpretability of the network.Data augmentation and focal loss are used to alleviate the impact of imperfect datasets.The experiments show that the proposed network still achieves good classification results with reduced parameters.(3)Propose a micronucleus cell image classification model that combines local and global attention mechanisms.The global attention mechanism complements the onesidedness of the single local attention and reduces the possibility of the network misclassifying dye pollution in cell images as micronuclei.The experiments show that the two attention mechanisms have good complementary advantages,and the fused attention not only significantly improves the interpretability of the network but also optimizes its classification performance,with AP,F1,and AUC values reaching 0.952,0.880,and 0.996,respectively.Therefore,the proposed method can effectively diagnose cell micronuclei and play an auxiliary role in clinical diagnosis.