Sharma Rohit, Kumar Mahanti Gautam, Panda Ganapati, Singh Abhishek
Department of Electronics and Communication Engineering, National Institute of Technology Durgapur, West Bengal, India.
Department of Electronics and Telecommunication, C. V. Raman Global University, Bhubaneswar, Orissa, India.
Curr Med Imaging. 2023 Apr 5. doi: 10.2174/1573405620666230405085358.
Thyroid disorders are prevalent worldwide and impact many people. The abnormal growth of cells in the thyroid gland region is very common and even found in healthy people. These abnormal cells can be cancerous or non-cancerous, so early detection of this disease is the only solution for minimizing the death rate or maximizing a patient's survival rate. Traditional techniques to detect cancerous nodules are complex and time-consuming; hence, several imaging algorithms are used to detect the malignant status of thyroid nodules timely.
This research aims to develop computer-aided diagnosis tools for malignant thyroid nodule detection using ultrasound images. This tool will be helpful for doctors and radiologists in the rapid detection of thyroid cancer at its early stages. The individual machine learning models are inferior to medical datasets because the size of medical image datasets is tiny, and there is a vast class imbalance problem. These problems lead to overfitting; hence, accuracy is very poor on the test dataset.
This research proposes ensemble learning models that achieve higher accuracy than individual models. The objective is to design different ensemble models and then utilize benchmarking techniques to select the best model among all trained models.
This research investigates four recently developed image transformer and mixer models for thyroid detection. The weighted average ensemble models are introduced, and model weights are optimized using the hunger games search (HGS) optimization algorithm. The recently developed distance correlation CRITIC (D-CRITIC) based TOPSIS method is utilized to rank the models.
Based on the TOPSIS score, the best model for an 80:20 split is the gMLP+ViT model, which achieved an accuracy of 89.70%, whereas using a 70:30 data split, the gMLP+FNet+Mixer-MLP has achieved the highest accuracy of 82.18% on the publicly available thyroid dataset.
This study shows that the proposed ensemble models have better thyroid detection capabilities than individual base models for the imbalanced thyroid ultrasound dataset.
甲状腺疾病在全球范围内普遍存在，影响着许多人。甲状腺区域细胞的异常生长非常常见，甚至在健康人群中也能发现。这些异常细胞可能是癌性的或非癌性的，因此早期发现这种疾病是降低死亡率或提高患者生存率的唯一解决办法。传统的检测癌性结节的技术复杂且耗时；因此，使用了几种成像算法来及时检测甲状腺结节的恶性状态。
本研究旨在开发用于使用超声图像检测恶性甲状腺结节的计算机辅助诊断工具。该工具将有助于医生和放射科医生在早期快速检测甲状腺癌。由于医学图像数据集的规模很小且存在严重的类别不平衡问题，单个机器学习模型不如医学数据集。这些问题导致过拟合；因此，在测试数据集上的准确率非常低。
本研究提出了比单个模型具有更高准确率的集成学习模型。目标是设计不同的集成模型，然后利用基准测试技术在所有训练模型中选择最佳模型。
本研究调查了四种最近开发的用于甲状腺检测的图像变换器和混合器模型。引入了加权平均集成模型，并使用饥饿游戏搜索（HGS）优化算法对模型权重进行优化。利用最近开发的基于距离相关CRITIC（D-CRITIC）的TOPSIS方法对模型进行排名。
基于TOPSIS评分，对于80:20的分割，最佳模型是gMLP+ViT模型，其准确率达到89.70%；而对于70:30的数据分割，gMLP+FNet+Mixer-MLP在公开可用的甲状腺数据集上达到了最高准确率82.18%。
本研究表明，对于不平衡的甲状腺超声数据集，所提出的集成模型比单个基础模型具有更好的甲状腺检测能力。
