| The utilization of deep convolutional neural networks(DCNN)in a variety of areas,such as intelligent transportation and medical care,has been greatly advanced by the swift advancement of internet technology,exemplified by deep learning.Despite the need for large samples of labeled data to be trained by DCNN,in actual situations it is challenging to acquire such a large quantity to satisfy the model’s training needs.Sometimes there is the problem of labeling difficulties even if the data is obtained.In addition,the premise that the training and validation sets of a machine learning model must be independently and identically distributed is a difficult one to fulfill in practical applications,thus making it necessary for machine learning models to meet this requirement.The advantage of transfer learning,which applies the knowledge learned by the model from a certain task to tasks or problems that are not identical but relevant,can overcome the independent homogeneous distribution constraint.,This advantage can further encourage us to overcome the problem that DCNN may not be able to obtain large amounts of labeled data by transfer learning.Therefore,transfer learning by using DCNN important for transfer learning tasks.In this paper,an improved transfer learning model is proposed for a small sample dataset,and the specific work consists of the following points.(1)This paper discusses the correlation between DCNN and transfer learning,and systematically summarizes them from different perspectives,and summarizes them into a complete classification system.(2)For the case that both source and target domains are labeled but have different tasks,most existing methods use the inherent training method,which is modifying classifier layers combined with fine-tuning,and such methods ignore the content variability between source and target domains,and feature extraction is highly subjective.Therefore,a deep transfer learning technique,based on enhanced ResNet,is suggested to augment the model’s feature recognition capability by augmenting the amount of network layers.The algorithm provides a useful combination and modification of the transfer learning-based model construction and training methods to avoid the problem of reduced feature recognition in the target domain due to the content discrepancy of the dataset and improve the recognition rate.Experiments on two datasets have been conducted specifically,and the outcomes demonstrate that this algorithm is superior in performance when compared to the comparison algorithm.Experiments on two datasets have been conducted specifically,and the outcomes demonstrate that this algorithm is superior in performance when compared to the comparison algorithm.(3)For the differentiation of lung adenocarcinoma patient subtypes,the histological phenotype identification of lung adenocarcinoma patients is crucial for treatment.In this paper,we present a feature-based transfer learning algorithm to tackle the problem,and in our retrospective study,we introduce a reliable transfer learning prediction model.This prediction model predicts the EGFR mutation status of lung adenocarcinoma patients based on noninvasive CT images with the clinical characteristics of the patients.This study shows that there is a clear association between higher-order features of CT images and EGFR genotype,and the feature-based transfer learning algorithm has the ability to identify EGFR mutation status.Therefore,a non-invasive based approach to predict EGFR mutations is possible before using invasive biopsies with expensive molecular assays.This study confirms the value of transfer learning for applications in the field of smart medicine. |
