【打印本页】 【下载PDF全文】 【HTML】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 326次   下载 277 本文二维码信息
码上扫一扫!
双模型策略在指甲病图像智能诊断中的应用
陈军晓1,殷杰1,扈东营2,吴兆1,朱秀艳1,王士勇1*
0
(1. 海军军医大学(第二军医大学)第三附属医院信息科, 上海 200438;
2. 海军军医大学(第二军医大学)第二附属医院皮肤科, 上海 200003
*通信作者)
摘要:
目的 探索一种在小数据量条件下提高医学诊断神经网络模型准确率和泛化能力的方法,解决在指甲病图像计算机辅助诊断中由于训练数据量小而导致神经网络模型性能不佳的问题。方法 提出融合实例分割与细粒度特征分类的双模型策略,采用第一届全国数字健康创新应用大赛健康医疗大数据主题赛基于图像的指甲病智能诊断模型任务数据集训练基于双模型策略的神经网络模型,该任务数据集涵盖甲母痣、甲沟炎、银屑病甲、甲真菌病、甲下出血、甲黑线、甲周疣、甲黑素瘤8类指甲病,各类别不平衡。评估双模型策略的诊断性能,并与相同软、硬件训练条件下单模型策略[图像分类模型(ResNet50、Swin Transformer)和基于快速区域卷积神经网络(Faster R-CNN)的目标检测模型]进行比较。结果 纳入任务数据集包括甲母痣210例、甲沟炎186例、银屑病甲69例、甲真菌病203例、甲下出血149例、甲黑线71例、甲周疣93例、甲黑素瘤67例共1 048例样本,其中90%的样本用于训练不同策略的模型,10%用于评估模型。基于ResNet50的图像分类模型的micro F1值为0.324,基于Swin Transformer的图像分类模型为0.381,基于Faster R-CNN的目标检测模型为0.572,基于双模型策略的Mask R-CNN+Swin Transformer模型为0.714。双模型策略预测各指甲病的准确度为甲母痣80.95%(17/21)、甲沟炎89.47%(17/19)、银屑病甲100.00%(7/7)、甲真菌病70.00%(14/20)、甲下出血73.33%(11/15)、甲黑线14.29%(1/7)、甲周疣55.56%(5/9)、甲黑素瘤42.86%(3/7)。双模型策略在该任务1 000例测试集中的micro F1值为0.844。结论 双模型策略可以有效结合功能不同的模型,更好地完成小数据量训练条件下的指甲病图像智能诊断任务。
关键词:  指甲病  智能诊断  神经网络  实例分割  细粒度特征分类
DOI:10.16781/j.CN31-2187/R.20240166
投稿时间:2024-03-15修订日期:2024-06-27
基金项目:
Application of dual-model strategy in image intelligent diagnosis of nail diseases
CHEN Junxiao1,YIN Jie1,HU Dongying2,WU Zhao1,ZHU Xiuyan1,WANG Shiyong1*
(1. Department of Information, The Third Affiliated Hospital of Naval Medical University (Second Military Medical University), Shanghai 200438, China;
2. Department of Dermatology, The Second Affiliated Hospital of Naval Medical University (Second Military Medical University), Shanghai 200003, China
* Corresponding author)
Abstract:
Objective To explore a method to improve the accuracy and generalization ability of medical diagnostic neural network models under conditions of small data volumes, and to address the issue of poor neural network model performance in computer-aided diagnosis of nail diseases due to limited training data. Methods A dual-model strategy integrating instance segmentation with fine-grained feature classification was proposed. The neural network model based on dual-model strategy was trained using the dataset of Image-Based Intelligent Diagnosis of Nail Disease Model task of the first National Digital Health Innovation Application Competition & Health and Medical Big Data Theme Competition. This dataset covered 8 types of nail diseases, including nail matrix nevi, paronychia, nail psoriasis, onychomycosis, subungual hemorrhage, melanonychia, periungual warts, and nail melanoma, with class imbalance present. The diagnostic performance of the dual-model strategy was evaluated and compared with single-model strategies (image classification models [ResNet50 and Swin Transformer] and target detection model based on faster region-based convolutional neural network [Faster R-CNN]) under the same hardware and software training conditions. Results The dataset included 1 048 samples, including 210 cases of nail matrix nevi, 186 cases of paronychia, 69 cases of nail psoriasis, 203 cases of onychomycosis, 149 cases of subungual hemorrhage, 71 cases of melanonychia, 93 cases of periungual warts, and 67 cases of nail melanoma, with 90% used for training various models and 10% for evaluation. The micro F1 score was 0.324 in the image classification model based on ResNet50, 0.381 in the image classification model based on Swin Transformer, 0.572 in the target detection model based on Faster R-CNN, and 0.714 in the dual-model strategy model Mask R-CNN+Swin Transformer. The accuracy rates for diagnosing different nail diseases in the dual-model strategy were: nail matrix nevi 80.95% (17/21), paronychia 89.47% (17/19), nail psoriasis 100.00% (7/7), onychomycosis 70.00% (14/20), subungual hemorrhage 73.33% (11/15), melanonychia 14.29% (1/7), periungual warts 55.56% (5/9), and nail melanoma 42.86% (3/7). The micro F1 score for evaluating the dual-model strategy on a test set of 1 000 cases was 0.844. Conclusion The dual-model strategy can effectively combine models with different functions to well accomplish the task of intelligent diagnosis of nail diseases under small data volume training conditions.
Key words:  nail disease  intelligent diagnosis  neural network  instance segmentation  fine-grained feature classification