A systematic review on artificial intelligence techniques for detecting thyroid diseases

Probabilistic approaches

In Rao & Renuka (2020), Naive Bayes and a Decision Tree, built using the ID3 algorithm, are used to perform a binary prediction about whether the patient is affected by hypo- or hyperthyroidism. The authors of the study in Pasha & Mohamed (2020) perform feature selection on the UCI dataset for thyroid disease detection by exploiting both a Random Forest-based method and a Gain Ratio technique. In the end, the prediction is performed by comparing also different machine learning techniques, namely K-Nearest-Neighbor, Logistic Regression, and Naive Bayes. Similarly, Houari et al. (2016) try to reduce redundant dimensions by exploiting Copulas and LU-decomposition techniques. They test their methods also on the UCI thyroid dataset for detecting both hypo- and hyperthyroidism. For the evaluation of the data reduction techniques, the authors employed Naive Bayes, besides Artificial Neural Networks (ANN) and k-nearest neighbors (k-NN) as learning algorithms. The contribution in Chandel et al. (2016) compares different ML techniques in detecting thyroid-related diseases (i.e., hypo- and hyperthyroidism), with a particular focus on Naive Bayes, K-Nearest Neighbor, and Support Vector Machine. Juneja (2022) presents a fuzzy adaptive feature filtration and expansion-based model to generate a novel feature set related to thyroid. The obtained feature set is then analyzed through Extreme Learning Machine classifiers, whose performance is compared with Naïve Bayes, Decision Tree, Multilayer Perceptron, and Radial Basis Function networks. Kishor & Chakraborty (2021) compare seven machine learning classifiers such as decision tree, support vector machine, Naïve Bayes, adaptive boosting, Random Forest, artificial neural networks, and K-nearest neighbor. They find out that Random Forest is the best performing algorithm to predict fatal diseases about thyroid. Authors in Qin et al. (2021) try to study papillary thyroid cancer through the analysis of magnetic resonance imaging radiomics by exploiting eight classifiers (logistic regression, bagging, Random Forests, extremely randomized trees, support vector machines, Naïve Bayes, multilayer perception, and K-nearest neighbors). Some of the models succeeded into reaching a performance of correct classification higher than 95%.

In Peya, Chumki & Zaman (2021), a thyroid disease prediction model is proposed exploiting three machine learning classification algorithms (i.e., K-Nearest Neighbor, Naive Bayes, and Decision Trees). Using the thyroid data of the UCI machine learning repository and a 10-fold cross-validation, the decision tree resulted the most accurate algorithm, with a 99.7% of accuracy.

Riajuliislam, Rahim & Mahmud (2021) try to predict early stage hypothyroidism. They apply three different feature selection procedures: Principal Component Analysis (PCA), Recursive Feature Elimination (RFE), and Univariate Features Selection (UFS). Moreover, they use different classification algorithms (i.e., support vector machine, decision tree, random forest, logistic regression, and Naive Bayes). RFE applied to the UCI thyroid dataset allowed the authors to achieve a constant 99% accuracy value over the various considered classification algorithms.

In Raisinghani et al. (2019), the authors compare different machine learning approaches, (logistic regression, decision trees, random forest, support Vector Machine) to develop predictive models to detect a generic thyroid disease.

The contribution in Chandio et al. (2016) exploits time series regression to create an intelligent system for thyroid disease visualization. This allows a careful surveillance of the thyroid disease, with a particular focus on hypothyroidism, hyperthyroidism, and euthyroid disease.

The study in Chaubey et al. (2021) compares logistic regression, decision tree, and kNN on the UCI knowledge discovery database for thyroid diseases detection. The best result of accuracy is obtained by applying the kNN classifier (96.87%).

Finally, Islam et al. (2022) try to detect euthyroid disease by leveraging eleven machine learning classifiers, namely, Gaussian naive Bayes, ANN, CatBoost, XGBoost, Random Forest, LightGBM, Decision Tree, Extra-trees, SVC, and KNN, and applying also feature selection to the sick-euthyroid dataset. The best performing classifier resulted to be ANN with a maximum accuracy equal to 95.87%.

Kernel-based approaches

Duggal & Shukla (2020) perform feature selection and extraction before applying naive Bayes, Support Vector Machine, and Random Forest to identify hypothyroidism, hyperthyroidism, and euthyroid disease.

The authors of Shahid et al. (2019) compare Random Forest, Support Vector Machine, and K-Nearest Neighbours on the UCI thyroid dataset, to discover the best performing algorithm, resulted to be Random Forest, in detecting hypo- and hyperthyroidism.

The contribution in Ahmed & Soomrani (2016) provides a framework, named Thyroid Disease Types Diagnostics (TDTD), aiming at making a diagnosis of various thyroid diseases in a very structured and transparent manner and exploiting binary and multi-SVM algorithms, as well as Bayesian isotonic regression for missing values. In Tyagi, Mehra & Saxena (2018), the authors present again the results in detecting hypo- and hyperthyroidism (by using the UCI dataset) through different machine learning techniques, such as decision trees, artificial neural networks, support vector machines, and k-nearest neighbors.

Kumar (2020) introduces a novel mutliclass SVM approach to detect four types of subjects, i.e., people affected by hypothyroidism, hyperthyroidism, euthyroidism disease and healthy euthyroid, on the already mentioned UCI dataset. Shen et al. (2016) propose a novel scheme to optimize the parameters of SVM by means of the fly optimization algorithm. The novel scheme is compared with other optimization algorithms and its efficiency is tested on four datasets, including the UCI one about thyroid.

In Li et al. (2019b), a novel optimization technique for SVM applied to the UCI thyroid dataset is proposed. It is based on the teaching-learning algorithm and differential evolution and it permitted to SVM to reach better performance in comparison with other solutions.

Malignant nodules detection is the focus of the work in Raghavendra et al. (2018). The paper presents a computer-aided diagnosis system to detect thyroid malignant nodules by means of higher order spectral entropy features and using particle swarm optimization (PSO) and support vector machine (SVM) frameworks.

In Prochazka et al. (2019), the authors propose a computer-aided diagnosis system using direction independent features of ultrasound images of the thyroid gland, with the aim of detecting malignant nodules by means of Random Forest and SVM.

Similarly, the authors of the study in Raghavendra et al. (2017) try to propose an SVM-based computer-aided diagnosis system, by exploiting fusion of Spatial Gray Level Dependence Features (SGLDF) and fractal textures to detect benign and malignant thyroid lesions in ultrasound images.

Kaur, Kumar & Kumar (2019) propose an IoT-based framework using SVM, Random Forest, Decision Trees, k-nearest neighbour, and artificial neural networks to be tested on different healthcare datasets, including one about general thyroid disease.

Shen et al. (2021) apply SVM to platelet RNA-seq data in order to differentiate different types of thyroid cancer and are able to achieve an accuracy of 97%.

Decision Tree approaches

Hayashi (2017) tries to create a white-box model to make prediction on the UCI thyroid datasets by exploiting the synergical effects of Recursive-Rule eXtraction (Re-RX) with J48 graft in terms of rule definition in the IF-THEN form for predicting both hypo- and hyperthyroidism. In Sidiq & Mutahar Aaqib (2019), the authors employ decision trees, random forest, vote ensemble, and stack ensemble in order to detect both hypo- and hyperthyroidism. Hao et al. (2018) introduce a decision tree improved by MS-Apriori for the prognosis of lymph node metastasis (LNM) in patients with thyroid cancer. MS-Apriori is used to generate association rules considering rare items by multiple supports and fuzzy logic is introduced to map attribute values to different sub-intervals. The used dataset is made of clinical-pathological data, obtained from the First Hospital of Jilin University.

Jha et al. (2018) present a hybrid algorithm for healthcare data mining by using Decision Stump (DS), StackingC (SC), and voting methods to tackle the hyperthyroidism issue. In Imbus et al. (2017), a random tree and a rule-based classifier (JRip) are used to detect primary hyperparathyroidism; while in Pan et al. (2016), random forest, together with principal component analysis and rotation transformation, is used to detect a general thyroid disease by exploiting the UCI thyroid dataset.

Finally, Zarin Mousavi, Mohammadi Zanjireh & Oghbaie (2020) apply computational methods based on decision trees, like ID3 and CHAID, SVM and multi-layer perceptrons, enriched with bagging and boosting techniques, for the identification of congenital hypothyroidism.

Rule-based approaches

In Bentaiba-Lagrid et al. (2020), a new amplification technique, based on randomization for systems incorporating a structured case-base that speeds up case retrieval while supporting case retention, is presented. The case-base is segmented in a novel manner, with new similarity functions based on features’ weights in order to accelerate the retrieval of the case-base reasoning. The system was applied to the detection of both hypo- and hyperthyroidism and compared with different machine learning methods.

In Asaad Sajadi et al. (2019), hypothyroidism is detected by means of a fuzzy rule-based expert system, which is proved to perform better than a logistic regression model on a real dataset.

Finally, the authors of Kumari & Sharma (2019) draft a fuzzy logic-based expert system to tackle both hypo- and hyperthyroidism.

Neural networks

Mahurkar & Gaikwad (2017) exploit artificial neural networks in conjunction with K-Means to normalize raw data for hypo- and hyperthyroidism. In Vivar et al. (2020), a guiding computer-aided diagnosis system, using a neural network with a dropout at the input layer, and integrated gradients of the trained network at test-time to attribute feature importance dynamically, is proposed. The technique is applied also to the UCI thyroid dataset to detect both hypo- and hyperthyroidism.

Similarly, in Santos et al. (2019), the authors introduce a decision support system, to make general thyroid dysfunction assessment, based on an artificial neural network and complemented by a novel approach to knowledge representation and argumentation. In Yue et al. (2020), Fourier transform infrared spectroscopy was combined with three neural network models, namely multilayer perceptron, long-short-term memory network, and a convolutional neural network in order to detect hypo- and hyperthyroidism.

Chai (2020) tackles thyroid disease in general by means of knowledge graphs, by extracting the relationships between bio-medical entities for feeding a bidirectional long short-term memory network. This combination proved to have better diagnostic effects than other techniques on an image dataset from the university of Shanghai. In Zhang, Huang & Lv (2020), the authors propose a synthetic data augmentation method based on progressive generative adversarial network in order to improve the performance in deep learning detection of hypo- and hyperthyroidism.

Yin et al. (2019) propose a hybrid cutting network, featuring a regional attribute cutting method, for feature extraction and classification applied to a dataset of thyroid ultrasound images. The objective was to detect malignant thyroid nodules that could cause cancer.

In Yi et al. (2017), the authors propose a novel diagnostic system to detect thyroid cancer, with a particular focus on thyroid nodule risk assessment. The method employs convolutional neural networks to analyze ultrasound images.

The contribution in Lyu & Haque (2018) embeds high dimensional RNA-Sequence data into bi-dimensional images and uses a convolutional neural network to detect various types of cancer, including thyroid cancer.

In Li et al. (2019a), the authors employ deep convolutional neural networks to enhance the diagnostic accuracy of thyroid cancer through the analysis of sonographic images coming from clinical ultrasounds. Ultrasound images are also used in Poudel et al. (2019) to feed convolutional neural networks for texture classification of anatomical structures of the thyroid for detecting general changes of its shape. Conversely, in Moran et al. (2018), thermograms are exploited by convolutional neural networks for the early identification of thyroid nodules that can possibly cause cancers. Guo & Du (2019) exploit again ultrasound images of the standard plane of the thyroid to evaluate its general status by means of deep convolutional neural networks. In particular, a 18-layer ResNet achieves the best results according to the authors. In Shi et al. (2020), the authors integrate domain knowledge, extracted from standardized terminology, and deep learning (Auxiliary Classifier Generative Adversarial Network) into a synthetic medical image augmentation technique to classify ultrasonography thyroid nodules.

In Lu et al. (2020), hyperthyroidism is tackled, with a particular focus on its progression, by means of enhanced LSTMs with an adaptive loss function. The analyzed data regard blood test information in the early stage from a Shangai hospital.

Cordes et al. (2021) detect thyroid malignant nodules using artificial neural networks on ultrasonographic characteristics obtaining an accuracy of 84.4%. Similarly, Jin et al. (2021) analyze clinical ultrasound imaging data, in five hospitals in China, through artificial neural networks obtaining performance between 80% and 90%. Ahmed et al. (2022) use artificial deep neural networks on the concatenation of 6 databases containing data collected from the Garvan Institute in Sydney, Australia, to discover thyroid cancer. The main and only shown result is the accuracy of 98%; other results and the optimization process are not described in detail.

Ananthi et al. (2022) and Khan (2021) exploit convolutional neural networks to try to detect the onset of thyroid dysfunctions. The former employs X-ray images to prevent hypo- and hyperthyroidism as well as malignant nodules and other dysfunctions reaching a 99% of accurate prediction. The latter only focus on hypothyroidism and achieve 98% of accuracy. Liu et al. (2021) tackle the issue of malignant nodes identification using convolutional neural networks as well. In this case, they apply information fusion techniques on ultrasound images and radio frequency signals, achieving better results than using only ultrasound images in detecting malignant thyroid nodules.

In Santillan et al. (2021) and Song et al. (2022) the focus is on the detection of malignant thyroid nodules which could possibly generate cancer. The former apply convolutional and recurrent neural networks to Fourier Transform infrared spectroscopy data achieving accuracy of 98.06%. The latter employ a feature-enhanced dual branch convolutional neural network on ultrasound images of the thyroid gland, obtaining the best mean average precision of identification equal to 92.5%. Chu, Zheng & Zhou (2021) face the issue of malignant thyroid nodule detection too, by employing a mark-guided ultrasound deep network segmentation model which, in turn, is based on different types of convolutional neural networks. They achieve a segmentation accuracy equal to 97.85%, improving the outcomes of other standard convolutional neural networks.

Yang et al. (2021) use convolutional neural networks on thyroid scintigrams to detect general dysfunctions of the gland and report an accuracy of 92.73%. Similarly, Pi et al. (2022) exploit convolutional neural networks and the fusion of deep and handcrafted features from thyroid scintigraphy to early detect thyroid dysfunctions. They were able to reach an accuracy equal to 91.18% and an f-measure equal to 88.11%.

In Zhao et al. (2022), semantic consistency generative adversarial network are employed to detect malignant thyroid nodules using ultrasound data. The proposed method is claimed to reach an accuracy equal to 94.30% and an area under the curve equal to 97.02%.

Hosseinzadeh et al. (2021) use a multiple multilayer perceptron neural network to identify hypo- and hyperthyroidism in the context of Internet of Medical Things, reaching an accuracy of 99%. Saktheeswari & Balasubramanian (2021) exploit an autoencoder-based neural network using also a multi-layer tree-based state machine to detect malignant thyroid nodules as well as hypo- and hyperthyroidism, with a final mean accuracy equal to 98.90%.

Finally, Ai et al. (2022) apply a recent type of neural networks, i.e., capsule networks, on ultrasonic thyroid images to detect possible thyroid cancer traces achieving a top accuracy equal to 81.06%.

Ensembles

In Yadav & Pal (2022), an ensemble of different machine learning techniques is employed to detect thyroid hormone disease. The ensemble uses Boosting, Bagging, Stacking, and Voting and it is aimed at identifying hypothyroidism patients.

In Alam, Siddique & Adeli (2020), the authors present a novel dynamic ensemble learning of neural networks. It provides an automatic design of the ensemble, a maintaining of accuracy and diversity of the composing neural networks, and very few parameters to be designed by the user. The technique is successfully employed to detect hypothyroidism, hyperthyroidism, and euthyroid disease.

Priyadharsini & Sasikala (2022) exploit Adaboost and Bagging as ensemble machine learning methods to correctly detect hypo- and hyperthyroidism. Bagging resulted in better performance as regards all the main metrics: accuracy 99.20%, precision 99.9%, f-measure 99.8%.

Finally, Yadav & Pal (2022) exploit Boosting, Bagging, Stacking, and Voting ensembles to detect hypothyroidism achieving a top accuracy equal to 99.86% and a top recall equal to 99.88%.

Other learning approaches

In Priya & Manavalan (2018), a weighted extreme learning machine technique hybridized with Invasive Weed optimization is used to detect general thyroid disease. Conversely, in Pavya & Srinivasan (2017), filter-based and wrapper-based feature selection methods are applied to four classifiers, namely, MultiLayer Perceptron Back Propagation Neural Network, Support Vector Machine, and Extreme Learning Machine, in order to detect both hypo- and hyperthyroidism. In Ma et al. (2018), the authors introduce a novel hybrid diagnosis system, integrating local fisher discriminant analysis and kernelized extreme learning machine method for thyroid disease diagnosis (hypo- and hyperthyroidism). Finally, in Baccour (2018), a new classification system for both hypo- and hyperthyroidism and based on fused VIKOR and TOPSIS is proposed.