Histologic Images Analysis Of Mycosis Fungoides Based On Deep Neural Networks

Mycosis fungoides is a kind of skin tumor with a high degree of malignancy,which is easily confused with some benign skin inflammations in the early stage.Therefore,the accurate identification and diagnosis of mycosis fungoides is an urgent problem to be solved.Early diagnosis and early treatment of mycosis fungoides will greatly improve the cure rate of patients and thus benefit patients.Histologic diagnosis is currently the gold standard for the diagnosis of mycosis fungoides.Skin pathologists usually look for some effective features of mycosis fungoides by analyzing various tissue components and the related morphology of cells in the patient's skin histologic images.And making a diagnosis by analyzing these characteristics and the doctor's experience.However,due to the huge size of the histologic image,the variety of tissue types,the complex scenes,and the chaotic cell distribution,the work of the evaluation and diagnosis of the doctor under the microscope is extremely heavy.In order to improve the accuracy of diagnosis,there is an urgent need for some automatic and quantitative methods for histologic images of mycosis fungoides.Under this situation,this paper proposed a set of automatic segmentation models for multiple tissue components of mycosis fungoides,which could automatically segment multiple types of tissues in complex scenes.At the same time,this paper studied the instance segmentation network framework of the cell nucleus,so that the morphology and histology of the cell nucleus could be analyzed.In the segmentation of multiple tissues in the pathological image of mycosis fungoides,this paper proposed a cascade deep network model.First,this paper proposed a Fast and Dense VGG(FD-VGG)for faster initial segmentation of multiple organizations.The network used VGG as a backbone network,used convolution without padding,and changed the final pooling layer to average pooling,so that the network could be trained with small image patches and directly tested with large image patches,which greatly reduced the segmentation time.The post-processing algorithm was used to remove some errors of the initial segmentation,and then the final multiple tissue segmentation results were obtained through the boundary optimization model Segnet.The accuracy of the segmentation model on the validation set was 0.953,and it took only 15 seconds to test a pathological image containing approximately 300 million pixels.The qualitative and quantitative comparison between the model and the network with padding proved that the segmentation of the network model proposed in this paper is more accurate.In the segmentation task of the cell nucleus,this paper presented a novel instance segmentation network of the cell nucleus(Horizontal and Vertical SENet,HV-SENet)by analyzing the deficiencies of the currently used semantic segmentation method in the cell nucleus instance segmentation task.The network used multi-task learning to combine the semantic segmentation of the nucleus and the learning of horizontal-vertical maps.The two shared an encoder and learn collaboratively,thus training a more powerful encoder.Finally,the post-processing algorithm was used to obtain the final segmentation result of the nuclear instance.The average DICE coefficient on the verification set of the nuclear segmentation model proposed in this paper reached 0.802,the DQ value was 0.560,the SQ value was 0.744,and the PQ value was 0.418.Comparison with other models proved that the model performance of this paper was better.Finally,it was tested in the whole slide image of mycosis fungoides,and satisfactory results were obtained.Therefore,the model in this paper could better segment various tissues and cell nuclei of mycosis fungoides,so as to assist skin pathologists to diagnose mycosis fungoides.