1.College of Computer Science and Technology (College of Data Science),Taiyuan University of Technology,Taiyuan,Shanxi,China
2.School of Software,North University of China,Taiyuan,Shanxi,China
3.School of Software,Taiyuan University of Technology,Taiyuan,Shanxi,China
4.Shanxi Radiation Therapy (Interventional Radiology) Clinical Medical Research Center,First Hospital of Shanxi Medical University,Taiyuan,Shanxi,China
Purposes Aiming at the problem of sample category imbalance in clinical application scenarios of medical images, a dual-branch feature fusion network with mixed cross-attention as the core is proposed. Methods The input image was subjected to feature fusion and importance score at the input end of the network, which promotes the model to focus on extracting the region of interest and alleviates the imbalance between the foreground and background of the image. An improved focal loss function for the two-classification task was proposed. The centrosymmetric characteristics of cubic power function was used in the proposed loss function. The penalty term of the misclassification was smoothed by improving the focus loss weighting factor and the introduction of fewer hyperparameters. Then the class imbalance factor of the focus loss was theoretically derived. The reference range was given to reduce the parameter search space. Second, aiming at the problem of insufficient sample size, the data enhancement method and sample replacement strategy for CT images were adopted to effectively increase the number of available samples. Results The experimental results on public data sets and real medical clinical datasets show that the improved focus loss improves the stability of the classification model. The proposed method can achieve more than 82% accuracy in the preoperative screening task of patients with portal hypertension, and also steadily improve task performance in multiple public datasets.
类别不平衡问题的存在使得传统损失函数,如交叉熵损失等无法有效约束训练模型以识别所有类别,倾向于更多关注多数类样本。为解决这一问题,研究者通常会在损失函数中引入一系列超参数,以实现对难易样本以及多数样本和少数样本的不同效果是解决问题的主要思路之一[8]。在目标检测与分割领域,前景和背景之间的严重不平衡影响了检测器的性能。KAIMIN[9]团队率先引入焦点损失来提高探测器的性能,取得了非常好的结果。该方法通过为具有挑战性或容易被错误分类的样本分配更高的权重,同时降低简单样本的权重来解决类不平衡问题。RIDNIK et al[10]将这种方法扩展到多标签分类任务,设计了一种新的不对称损失为正负样本分配不同的权重,并采用不同的损失计算机制。文献[6]从训练过程入手,提出了一种新的多加权损失函数和端到端累积学习策略,降低了异常样本对训练的影响。此外,WANG et al[10]将质量因素嵌入到焦点损失中,使网络能够专注于学习高质量的样本,而不是难以分类的样本。以上方法增强了焦点损失的灵活性,提升了性能表现,但都不同程度地引入额外的超参数,显著扩大了参数搜索空间。本文通过引入非参数的幂函数优化焦点损失,仅保留类别因子一项超参数,从源头降低参数搜索工作的负担,此外本文通过理论和实验推导给出唯一超参数的搜索范围与搜索方法,进一步降低所提方法的应用成本。
RAZZAKM I,NAZ S, ZAIBA.Deep learning for medical image processing: Overview,challenges and the future[J].Classification in BioApps:Automation of decision making,2017:323-350.
WUJ, FUR, FANGH,et al.Medsegdiff:Medical image segmentation with diffusion probabilistic model[C]//Medical Imaging with Deep Learning.PMLR,2024:1623-1639.
ZHAOZ J, RENX T, SONGK,et al.Traditional Chinese medicine prescription generation model based on search enhancement[J].Journal of Taiyuan University of Technology,2025,56(1):114-126.
LIQ, LIR R, QIANGY,et al.Researsh and progress of artificial intelligence in medical CT image reconstruction[J].Journal of Taiyuan University of Technology,2023,54(1):1-16.
[8]
YAOP, SHENS, XUM,et al.Single model deep learning on imbalanced small datasets for skin lesion classification[J].IEEE transactions on medical imaging,2022,41(5):1242-1254.
[9]
AZAMM A, K B SALAHUDDINSKHAN,et al.A review on multimodal medical image fusion:Compendious analysis of medical modalities,multimodal databases,fusion techniques and quality metrics[J].Computers in Biology and Medicine,2022,144:105253.
JOSHIK, KUMARM, TRIPATHIA,et al.Latest trends in multi-modality medical image fusion:A generic review[J].Rising Threats in Expert Applications and Solutions,2022,2022:663-671.
[15]
DE FRANCHISR, BOSCHJ, GARCIA-TSAOG,et al.Baveno VII–renewing consensus in portal hypertension[J].Journal of hepatology,2022,76(4):959-974.
[16]
LINX, GAOF, WUX,et al.Efficacy of albumin–bilirubin score to predict hepatic encephalopathy in patients underwent transjugular intrahepatic portosystemic shunt[J].European Journal of Gastroenterology & Hepatology,2021,33(6):862-871.
[17]
GUPTAA, GUPTAR.ISBI 2019 C-NMC challenge:Classification in cancer cell imaging.Select Proceedings[J].ISBI 2019 C-NMC Challenge:Classification in Cancer Cell Imaging,2019.DOI:10.1007/978-98115-07984 .
[18]
CHENX, WANGT, JIZ,et al.3D automatic liver and spleen assessment in predicting overt hepatic encephalopathy before TIPS:a multi-center study[J].Hepatology International,2023,17(6):1545-1556.
[19]
CAIW, LINH, QIR,et al.Psoas muscle density predicts occurrences of hepatic encephalopathy in patients receiving transjugular intrahepatic portosystemic shunts within 1 year[J].CardioVascular and Interventional Radiology,2022,45(1):93-101.