| In disease diagnosis,pathological image diagnosis,as the final diagnosis basis for most diseases,is one of the most accurate diagnosis methods and is called the gold standard for disease diagnosis.However,due to the relatively lagging development of pathology in China,problems such as a large shortage of pathologists,unbalanced distribution of pathological resources,and strong subjectivity in pathological diagnosis are common in China.With the development of artificial intelligence,advanced artificial intelligence technologies such as deep learning have shown a level close to or surpassing that of human experts in many fields such as image recognition and medical image diagnosis.It is possible to solve the above problems through artificial intelligence-assisted diagnosis technology.In particular,in the intelligent diagnosis of pathological skin images,many scholars have devoted themselves to the research of related technologies and have achieved certain results.Due to the diversity of pathological skin tissues,the complexity of cell morphology and structure,and the black-box features of neural networks,the accuracy,efficiency and interpretability of existing skin pathological image intelligent diagnosis techniques still need to be improved.Because of the problems mentioned above in intelligent diagnosis of pathological skin images,the thesis has carried out researches on large-scale pathological image data preprocessing,interpretable skin pathological image diagnosis,multi-scale pathological image feature extraction,and melanoma metastasis biomarker extraction.The work and major innovations achieved are as follows:(1)A data-parallel-based pathological image preprocessing method is proposed.Through the data-parallel method,a two-dimensional division method of the pathological image area is designed,the pathological image is divided into blocks,and each block is assigned to a processor core.A parallel algorithm for image block generation,image block screening,and parallel image block standardization are designed for segmented pathological image data.The original preprocessing is optimized by removing the dependencies and loop unrolling,which realizes the rapid preprocessing of pathological images.Experimental results show that the data-parallel-based pathological image preprocessing method achieves a high speedup ratio,and the method achieves a performance speedup of 22.61 times at 32 threads compared to the serial implementation on the CPU.(2)An interpretable skin pathology diagnosis method based on deep learning is proposed.This method designs a class-normalized activation mapping algorithm based on class activation mapping.The generation process of the class discriminative localization map is optimized by using extreme value truncation and normalization to accurately display the accurate response of the model to the input image to solve the artefact problem of the gradient-weighted class activation mapping algorithm.In the diagnosis stage of pathological images,a counting method for determining the weight of the predicted value is designed.A weighting factor of the predicted value is introduced to highlight the weight of the image block with a more significant predicted value and improve the diagnostic accuracy of the pathological image.At the same time,a probabilistic heatmap is introduced to locate the lesion area in pathological images.The experimental results show that the proposed method reaches the diagnostic level of pathologists in the melanoma identification task and has higher diagnostic accuracy than the mean and counting methods.At the same time,this method can more accurately display the accurate response of the neural network,visualize the more discriminative features learned by the neural network,and assist doctors in understanding the inherent decision logic of the neural network.(3)A skin pathology image classification method is proposed based on a multi-scale neural network.This method designs a new multi-scale input neural network structure called alterable multi-scale input convolutional neural network.The alterable multi-scale neural network comprises a backbone network and multi-scale input branches inserted into the backbone network in concatenation.The input of multiple scale images enables the network to obtain multi-scale feature extraction capabilities.Multi-scale neural networks with different scale combinations have different recognition capabilities.In order to obtain a multi-scale neural network structure with strong recognition ability,two multi-scale network search algorithms are designed.The minimum cost-based search algorithm tests all input branches in turn by specifying the priorities of insert position of input branches of different scales,discarding input branches that reduce network performance,and selecting favourable input branches to insert into the network.The search algorithm based on the hill-climbing method uses all insert position of input branches to form a search space.Each time,the input branch with highest performance is selected from the current search space and inserted into the network,and the input branch that reduces the network performance is discarded until the input branch cannot improve the network performance or the search space is null.The experimental results show that the variable multi-scale neural network has higher classification accuracy than the original network.The two search algorithms can find an excellent multi-scale network structure at a lower search cost than the exhaustive search algorithm.(4)A method for diagnosing melanoma metastasis based on a multi-modal model was proposed.This method designs a biomarker mining network based on the attention mechanism.The weighted fusion of image block features using the attention mechanism makes the model more inclined to focus on image blocks with significant features.The Breslow’s depth is introduced as a predictor,and the prediction accuracy of the model is improved.The Breslow’s depth and patient information were encoded by one-hot encoding,and image features were concatenated and classified using a fully connected layer-based classifier.The experimental results show that compared with the single-modal model,the multi-modal model can more accurately predict whether melanoma metastases and mine the potential biomarkers related to metastasis in pathological images,which has essential guiding value for selecting clinical treatment options for melanoma. |
