Table 1.

A summary of related automated systems for facial disease diagnosis along with their limitations.

Article	Dataset	Abnormality	Method	Results	Limitation
Kong et al., 2018 20	1123 Patients (private)	Acromegaly	Open CV for face detection. Facial locations landmarks as features. Fronatization. SVM, LR, K-NN, CNN, and RF ensemble classifiers.	Precision = 96% Sensitivity = 96% Specificity = 96%	Detect only one type of disease (binary classification). Used manual segmentation. Detect the face for diagnosis (controlled face diagnosis). Utilized only Facial location landmarks as features. Used only spatial features. Used only handcrafted features. Did not use DL features.
Schneider et al. 21	117 Patients (private)	Acromegaly	Geometric and Gabor filter feature extraction methods. FIDA (facial image diagnostic aid) software	Accuracy = 81.9%	Detect only one type of disease (binary classification). Used only handcrafted features. Did not use DL features. Controlled face diagnosis. Very small dataset. Private dataset. Low accuracy. Used commercial software to perform diagnosis. Large feature space.
Meng et al. 22	124 Patients (private)	Acromegaly	35 Anatomical facial landmarks, 55 Angular, index, and linear features, Geometric features. LDA classifier	Accuracy = 92.86%	Detect only one type of disease (binary classification). Used only handcrafted features. Did not use DL features. Controlled face diagnosis. Very small dataset. Private dataset. Large feature space.
Zhao et al. 23	48 Patients (private)	Down syndrome	Geometric, contourlet transform, and LBP features. SVM classifier	Accuracy = 97.9% Precision = 100% Sensitivity = 95.8%	Detect only one type of disease (binary classification). Used only handcrafted features. Did not use DL features. Controlled face diagnosis. Very small dataset. Private dataset. Large feature space.
Zhao et al. 24	100 Patients (private)	Down syndrome	HCLM, geometric, and textural features ICA SVM classifier	Accuracy = 95.6% Precision = 95.3% Sensitivity = 95.3%	Detect only one type of disease (binary classification). Used only handcrafted features. Did not use DL features. Controlled face diagnosis. Very small dataset. Private dataset.
Zhao et al. 25	130 Patients (private)	Down syndrome	HCLM, geometric, Gabor, and LBP features ICA LDA classifier	Accuracy = 96.7%	Detect only one type of disease (binary classification). Used only handcrafted features. Did not use DL features. Controlled face diagnosis. Very small dataset. Private dataset.
	24 Patients (private)	14 Dysmorphic syndromes	HCLM, geometric, Gabor, and LBP features ICA SVM classifier	Accuracy = 97%	Used only handcrafted features. Did not use DL features. Controlled face diagnosis. Very small dataset. Private dataset. Large feature space.
Lui et al. 26	87 Images (private)	Autism	K means and histogram features. SVM classifier	Accuracy = 88.51% Sensitivity = 93.1% Specificity = 86.1%	Detect only one type of disease (binary classification). Used only handcrafted features. Used only spatial features. Did not use DL features. Controlled face diagnosis. Very small dataset. Private dataset. Relatively low accuracy.
Sajid et al. 27	2000 Images	Palsy	GAN for augmentation. VGG-16 and CNN for feature extraction. CNN for classification	Accuracy = 92.6% Sensitivity = 93.14% Precision = 92.91%	Grade only one type of disease. Used only spatial features. Controlled face diagnosis. Used individual feature extraction to perform classification. Utilized individual classifiers to perform classification. Large feature space.
Guo et al. 28	1840 Images (private)	UPFP	Dlib DAN	AUC = 60.66%	Detect only one type of disease (binary classification). Used only spatial features. Controlled face diagnosis. Private dataset. Low performance. Large feature space.
Kuan Wang and Jiebo Luo 29	8509 Images	20 Diseases	Manually segmented symptoms. Used binary features Employed color features, Hough transform, k means clustering	Accuracy = 80.2% Sensitivity = 77.2% Precision = 82.1%	Used manual segmentation to crop the abnormality. Used only handcrafted features. Did not use DL features. Controlled face diagnosis. Very small dataset. Relatively low performance. Unbalanced dataset. Large feature space.
Gurovich et al. 30	26,692 Images (private)	216 Syndromes	DeepGestalt model (holistic + local + CNN features extraction methods) CNN for classification	Top-10-accuracy = 91%	Utilized individual classifiers to perform classification. Public only for health professionals. Private dataset.
Pantel et al. 31	646 Images	17 Syndromes and normal faces (binary classification)	DeepGestalt model	AUC = 89%	Utilized individual classifiers to perform classification. Public only for health professionals. Private dataset. Performed binary classification.
Jin et al. 16	350 Images (public)	4 Diseases	Open CV (HOG + SVM) AlexNet, ResNet-50, VGG-16 SVM	Accuracy = 93.3%	Used only spatial features. Utilized individual spatial DL features. Utilized individual classifiers. Employed a large number of features to perform classification.
		Beta-thalassemia		Accuracy = 95% Sensitivity = 100% Specificity = 90% Precision = 90.9%

Note. SVM: support vector machine; LR: logistic regression; K-NN: k-nearest neighbor; CNN: convolutional neural networks; LDA: linear discriminate analysis; HCLM: Hierarchical Constrained Local Model; ICA: independent component analysis; LBP: local binary pattern; GAN: generative adversarial network; DAN: deep alignment network.