Automatic Diagnosis Of Cutaneous Mycosis Fungoides Based On Histomorphology

The diagnosis of cutaneous mycosis fungoides can be easily confused with skin inflammation in clinical practice,thus making patients miss the best treatment period.In this study,three-level histomorphological features of tissue,patch and cell were used to extract and establish a random forest model for automatic diagnosis of skin mycosis fungoides.First,tissue segmentation of whole slide pathological images of the skin is performed and probabilistic heatmaps of epidermal and hair follicle tissue are obtained.In this paper,SERes2Net50 is used to classify five tissues and backgrounds of epidermis,dermis,hair follicles,sweat glands,and fat.In the training stage,the slices are firstly sliced according to the labeling results,and the sliding window patches(224×224×3)are used to construct the training set and the validation set.The data size is 89320 pieces and 8497 pieces respectively.The model is trained using multi class cross entropy as the loss function,and the average accuracy metric is defined for evaluation in comparative experiments.Experiments show that SERes2Net50 achieves an average classification accuracy of 0.96 on the validation set.In the testing phase,blocks are also taken according to the sliding window(224×224×3),and the tested slices are sent to the network by patch for category prediction,and the segmentation results of the slices are finally obtained by splicing according to the coordinates of the image patches.Finally,the segmentation result is subjected to post-processing operations to obtain a probability heat map showing the epidermis-hair follicle tissue with high probability.Second,semantic segmentation of epidermal lymphoid cells and epithelial cells is performed.In this paper,a multi-layer Transformer encoding network(Multi Trans E)is proposed for the segmentation task of epithelial and lymphoid cells.The overall model is an encoding-decoding structure.In the encoding layer,a stack of 20-layer Transformer modules is used as an encoder to extract high-level semantic features of cells.In the decoding stage,a transposed convolution is performed with a fixed small stride of 2 to decode the features.In this paper,576(256 × 256 × 3)pathological image patches of mycosis fungoides epidermal tissue and 2203 Pan Nuke public data were mixed as the data set for the segmentation task,which was divided into training set and validation set according to 8:2,and finally the data was divided into training set and validation set.The set is trained and validated on Multi Trans E.Experiments show that the Dice coefficients of epithelial cells and lymphocytes are 0.78 and0.83,respectively.And the average Dice coefficients in comparison experiments exceed other baseline networks such as UNet++ and Vi T.Finally,the morphological features of the three scales of tissue-patch-cell are fused and a random forest classifier is constructed for automatic diagnosis.The data set in this task has a total of 119 pathological sections of skin mycosis fungoides and lichen planus,of which 77 are skin mycosis fungoides and 42 are lichen planus.There are three steps in automatic diagnosis.The first step is the extraction of morphological features at three scales.This paper first extracts tissue-level features based on the probability heat map of pathological slices obtained in the first task,and then proposes an adaptive cell graph algorithm(ACG)to construct a cell graph map and extract map features,and simultaneously calculate clinical features(area,area ratio)of the two types of cells.The graph features and the clinical features of cells are used as cell-level features.Finally,1000 image patches at the junction of the epidermis and dermis are taken for manual feature extraction such as color and texture,and the expected value of the feature is taken as the feature of the patch level.The second step is feature engineering.The optimal feature subset is obtained by selecting,reducing dimensionality and removing redundancy of the features of the three levels.The third step is to build an automatic diagnostic model.In this paper,a random forest model classifier is established,and training and testing are performed on pathological slices,in which the data is divided into training set and test set according to 6:4.Experiments show that the model has an AUC metric of 0.94,the performance results of individual features over three levels and the performance results of its ablation experiments.