Research On Cell Counting Of Pathological Images Based On Deep Learning

Pathology Images is one of the important medical image type.Compare to other medical image type like CT and Ultrasound,Pathology Images could provide cell-level information,which makes the diagnosis more reliable.It plays a critical role in the diagnosis process of many malignant diseases.However,manual assessment of pathology images is timeconsuming and tedious.Thus,the use of computer technology to automate the diagnosis and analysis of pathology images is particularly important.Digital pathology is a subject that specializes in the digitalization and automatic analysis of pathology images.The main topic of this research is the cell counting task of histopathology images.The cell counting task is part of the quantitative analysis of pathological images,and the precise cell counts plays a key role in both clinical diagnosis and medical research.The common challenges of pathology images analysis are data scarcity,category diversity and high computational cost.Current research about cell detection and localization has got some achievement but still need to handle many problems.There are large space for further study to improve the cell counting performance in precision,speed and generalization capability.To solve the problems mentioned above,and fully utilize the potential of deep learning technology in the medical image processing,we do the following works: Firstly,we promote a novel light-weighted convolutional network named LC-Net.This model translate the cell counting task to the structural regression task by using deep segmentation methods.LC-Net is inspired by the idea from Unet that fusing different level feature,and combines it with the multi-scale convolutional block,which makes the model be more concise and efficient.Compare to the mainstream deep learning based methods,our model provide more precise counting result with higher speed and less parameters.Then,we address the high false positive rate of the model by designing a loss function called color_penalty.Color_penalty uses traditional image processing technology and the color information to suppress the response in the fake foreground part of the image,which forces model to learn the morphological feature of cell and alleviates the overfit problem of the model.Finally,we conduct many experiments on several public benchmarks and our own dataset.LC-Net gets the average counts error 4.9 ± 1.2 on the MBM dataset and the corresponding harmonic mean is 0.86.Our method also achieve the harmonic mean of 0.81 on the CRC dataset.We also run the model on our unlabeled liver cancer dataset,the output of our model is close to the average human annotations.As showed by experiments,the method proposed by this paper can provide reliable results on different benchmarks,with competitive cell counts result and high computational efficiency.In general,we provide an efficient and robust solution to cell counting task.It shows the possibility of real clinical application and shares many idea for further researches.