Identifying congenital generalized lipodystrophy using deep learning- DEEPLIPO

doi:10.21203/rs.3.rs-1649906/v1

Congenital Generalized Lipodystrophy (CGL) is a rare autosomal recessive disease characterized by near complete absence of functional adipose tissue from birth. CGL diagnosis can be based on clinical data including acromegaloid features, acanthosis nigricans, reduction of total body fat, muscular hypertrophy, and protrusion of the umbilical scar. The identification and knowledge of CGL by the health care professionals is crucial once it is associated with severe and precocious cardiometabolic complications and poor outcome. Image processing by deep learning algorithms have been implemented in medicine and the application into routine clinical practice is feasible. Therefore, the aim of this study was to identify congenital generalized lipodystrophy phenotype using deep learning. A deep learning approach model using convolutional neural network was presented as a detailed experiment with evaluation steps undertaken to test the effectiveness. These experiments were based on CGL patient’s photography database. The dataset consists of two main categories (training and testing) and three subcategories containing photos of patients with CGL, individuals with malnutrition and eutrophic individuals with athletic build. A total of 337 images of individuals of different ages, children and adults were carefully chosen from internet open access database and photographic records of stored images of medical records of a reference center for inherited lipodystrophies. For validation, the dataset was partitioned into four parts, keeping the same proportion of the three subcategories in each part. The 4-fold cross-validation technique was applied, using 75% (3 parts) of the data as training and 25% (1 part) as a test. Following the technique, 4 tests were performed, changing the parts that were used as training and testing until each part was used exactly once as validation data. As a result, a mean accuracy, sensitivity, and specificity were obtained with values of [90.85% ± 2.20%], [90.63% ± 3.53%] and [91.41% ± 1.10%], respectively. In conclusion, this study presented for the first time a deep learning model able to identify congenital generalized lipodystrophy phenotype with excellent accuracy, sensitivity and specificity, possibly being a strategic tool for detecting this disease.

Congenital Generalized Lipodystrophy

Screening

Diagnosis

Phenotype

Deep learning

Congenital Generalized Lipodystrophy (CGL) is a rare autosomal recessive disease characterized by near complete absence of functional adipose tissue from birth. That reduction of energy deposition culminates to ectopic lipid accumulation in tissues such as muscle, liver, heart and arterial wall. Consequently, there is a severe and early insulin resistance, diabetes mellitus, hepatic steatosis and premature atherosclerotic disease, which may lead to early death [1, 2].

CGL diagnosis can be based on clinical data including acromegaloid features, acanthosis nigricans, reduction of total body fat, muscular hypertrophy and protrusion of the umbilical scar. Also, laboratory data and imaging tests can show other important health information. The identification and knowledge of CGL by the health care professionals is crucial once it is associated with severe and precocious cardiometabolic complications and poor outcome. Although CGL patient phenotype be quite characteristic, the rarity of this disease and its misdiagnosis with common conditions like undernutrition and athletic shape make difficult an early detection of cases, which may significantly improve the prognosis of these patients [3].

Image processing may be crucial in phenotyping, diagnosis, and even the identification of rare new diseases [5]. Artificial intelligence (AI), mainly through machine learning, provides algorithms capable of learning from data. Images are one of the types of data that AI, namely deep learning, is more fruitful at analyzing [5]. Convolutional neural networks (CNNs) automatically detect patterns of interest in images and have demonstrated image classification performance above the level of humans, [4]. Deep learning algorithms have been implemented in medicine and the application into routine clinical practice is feasible [6], thus it is an interesting strategy to assist the health professional in the diagnosis for the proper management of patients with rare diseases. Therefore, the aim of this study was to identify congenital generalized lipodystrophy phenotype using deep learning.

A deep learning approach model was presented as a detailed experiment with evaluation steps undertaken to test the effectiveness. These experiments were based on CGL patient’s photography database. To build and train the convolutional neural network model, it was used Python 3 and some libraries to help, such as Numpy v1.17.4 and Tensorflow v1.15. All experiments were run on a standard PC without a GPU card and a i5-4210u processor.

This study was performed in accordance with the Declaration of Helsinki and was approved by the University Hospital Walter Cantídio Ethics Committee, Fortaleza, Ceara, Brazil (nº 5.364.464). All the patients and their families gave formal consent to participate in the study by signing the free informed consent form prior to their inclusion.

The dataset consists of two main categories (training and testing) and three subcategories containing photos of patients with CGL, individuals with malnutrition and eutrophic individuals with athletic build. A total of 337 images of individuals of different ages, children and adults were carefully chosen from internet open access database and photographic records of stored images of medical records of a reference center for inherited lipodystrophies. In the search for photographic records published on open access platforms, a literature review was carried out. The searches were carried out in the Lilacs, PubMed and Scielo databases. Descriptors and their combinations in Portuguese and English were used with Boolean operators: “Congenital Generalized Lipodystrophy” OR "Berardinelli-Seip Syndrome" AND “physiopathological mechanisms” OR “phonotype” OR “clinical characteristics”; “Malnutrition” AND “physiopathological mechanisms” OR “phonotype” OR “clinical characteristics”.

There was no standardization for the acquisition or selection of photographic images of the patients. The clinical history of the 22 patients followed up at the outpatient referral clinic, whose images were included in the analysis, was assessed through medical records.

Several data augmentation methods were employed to artificially increase the size and quality of the dataset. This process helps in solving overfitting problems and enhances the model’s generalization ability during training.

In order to carry out the data augmentation process, geometric transformation techniques were used. Some images were rotated and zoomed using angles arbitrarily chosen by the author. In total, 8 processes were chosen, 6 of which consisted of rotating 45, 90, 180º, -90, -50 or -45 degrees. And the other 2 consist of zooming the image and rotating 18º or 114º. At the end of the process, a database was obtained with a total of 896 images.

The architecture of the proposed CNN model consists of two major phases: the feature extraction and the classification (Fig. 1). However, before images are sent to the network, it is necessary to make transformations to the images. These consist of converting to grayscale, resizing and normalization. After carrying out these transformations, it is necessary to concatenate each transformed image in a large matrix, which is sent to the first layer of CNN. In this work, the feature extraction phase consists of using two pooling layers with a 2x2 filter and two convolutional layers with the ReLu function to increase the non-linearity of the output and each layer of the attribute extraction phase receives information from the previous layer, the data in the output layer, in turn, is passed as input to the subsequent layer. At the end of the first phase, and with the model already trained, the classification of patients with or without CGL is made by the fully connected network, and for this work we used three hidden layers with 1024 artificial neurons at each.

The hyperparameters used to configure the CNN are shown in Table 1. In this, it is possible to identify that the amounts of convolution and hidden layers are smaller than the amounts of neurons per layers. This motivation was due to the fact that the computational cost increases exponentially when increasing the number of layers. Another hyperparameter that can be highlighted is the activation function, which is present in all neurons, except those present in the output layer, which use the sigmoidal function.

Table 1

– Model Hyperparameters
Number of convolutional layers	2
Filter feature order	[2x2]
Number of pooling layers	2
Learning rate	0,001
Number of neurons per layer	1024
Turn-off neurons percentage	0,2
Activation Function	ReLu
Maximum Number of Epochs	600
Number of hidden layer	5

For validation, the dataset was partitioned into four parts, keeping the same proportion of the three subcategories in each part. The 4-fold cross-validation technique was applied, using 75% (3 parts) of the data as training and 25% (1 part) as a test. Following the technique, 4 tests were performed, changing the parts that were used as training and testing until each part was used exactly once as validation data (Fig. 2).

In addition to the images extracted from open access platforms, photographic records of 22 CGL patients followed in a referral clinic were used. The mean age was 14.6 ± 10.7 years, 68.2% (15/22) were younger than 18 years old, and 59% (13/22) were women. The mean age of participants in the group < 18 years was 8.3 ± 4.7 years, and in the group ≥ 18 years was 28.0 ± 6.2 years. The mean age at CGL diagnosis was 7.4 ± 11.1 years, ranging from one month to 38 years of age. All the patients had hypoleptinemia, 95.4% (21/22) had low HDL-c, 86.4% (19/22) had hypertriglyceridemia, 68.2% (15/22) diabetes, 50% (11/22) hepatic steatosis, 41% (9/22) insulin resistance, 41% (9/22) hypercholesterolemia, and 18.2% (4/22) hypertension. Genetic analysis was performed in only 45.4% (10/22) of the patients. AGPAT2 and BSCL2 gene mutations were identified in 40% (4/10) and 60% (6/10) of them, respectively.

After applying cross-validation, the results associated with each of the four subgroups were composed. Initially, confusion matrices were used to visualize the true positives (TP) and negatives (TN), as well as the false positives (FP) and negatives (FN). With these values, which are shown in Fig. 3, it is possible to raise the capacity that the model generated by the CNN network has in classifying patients with or without CGL. In Fig. 3, in all four subgroups, true positives and negatives have a predominance in relation to false positives and negatives, which is an initial indication showing the good generalization capacity of the chosen model.

With the data obtained from each confusion matrix, other indicators of the model's performance can be measured, such as accuracy, sensitivity, and specificity. The first is indicative of true positive and negative with respect to all cases evaluated. Sensitivity and specificity, which are illustrated respectively, are similar indicatives. The first is the model's ability to correctly predict when the patient does not have CGL, whereas with specificity it is able to correctly predict patients who have CGL.

The mean accuracy, specificity, and sensitivity were obtained with values of [90.85% ± 2.20%], [90.63% ± 3.53%] and [91.41% ± 1.10%], respectively. Regarding the ability to correctly identify patients with CGL, although the test with subset D did not obtain a result above 80%, on sensitivity, it still had a good classification capacity. In terms of specificity, all subsets obtained a rate above 90%, this is a result that coincides with the reality of the data set, since there is a certain level of control in the images of patients without LGC (Table 2). Finally, it is also noted that the computational cost measured in the training stage is also similar for each of the subsets, and this is a relatively low time compared to models with many hidden layers and neurons by layers.

Table 2

Results from k-fold CV.
Training	Test	Accuracy	Specificity	Sensitivity	Computational costs (minutes)
A, B, C	D	89.29%	93.75%	78.13%	25.40
A, B, D	C	92.86%	93.75%	90.63%	26.00
A, C, D	B	89.29%	88.13%	92.19%	25.10
B, C, D	A	92.41%	93.13%	90.63%	27.00
	Mean	90.85%	90.63%	91.41%	25.88
	Standard deviation	2.21%	3.53%	1.10%	0.8382

It is estimated that there are more than 7000 rare diseases worldwide, 80% of which are of genetic origin and approximately 75% affect children [7, 8]. Although individually rare, collectively, these diseases affect about 350 million people [9, 10]. The lack of knowledge about the true frequency of the disease and about specific diagnostic criteria, as well as the lack of official clinical guidelines and the small number of geographically dispersed patients, hinder the clinical diagnosis and recruitment of these patients for clinical research [11, 12].

Images, one of the types of data that artificial intelligence is more productive in analysis, can be used in phenotyping, diagnosis and even in the identification of rare new diseases [13]. This is the first study to present a deep learning model for the analysis of photographic image of CGL patients and identification of phenotypic characteristics.

CGL is a low prevalence condition, affecting 1: 10,000,000 live births, but it is believed that of every four existing cases, only one is reported. Between 300 and 500 patients affected by LGC have already been reported in the medical literature [14], with a greater concentration of cases in Lebanon, Brazil, Portugal, Scandinavia, as well as in families with African ancestors [15, 16].

The phenotype of the CGL patients is quite characteristic and appear early, in the first years of life, however the rarity of this disease and the lack of knowledge of health professionals make it difficult to identify the clinical manifestations. Many features of the syndrome can be identified through close observation during clinical evaluation. The scarcity of subcutaneous adipose tissue gives patients a characteristic muscular appearance (muscular pseudohypertrophy) with prominence of superficial subcutaneous veins. acromegalic facies, large hands and feet, acanthosis nigricans and umbilical protrusion can also be observed [14].

Other diseases or disorders in nutritional status may have similar clinical manifestations and physical characteristics. Thus, the development of technologies that involve machine learning can assist clinical evaluation, especially in conditions where remote assistance may be necessary.

LipoDDx® is a free mobile application for the identification of different subtypes of lipodystrophies, which is effective in approximately 80% of cases in this first validation process. This is the first app to allow identification of a set of heterogenous rare diseases, however, a sequence of responses is requested, with no photographic analysis as strategy for identifying diseases, as proposed in the present study [17]. Gurovich Y et al (2019) presented a facial analysis framework for genetic syndrome classification called DeepGestalt [18]. This framework leverages deeplearning technology and learns facial representation from a largescale face-recognition dataset, followed by knowledge transfer to the genetic syndrome domain through fine-tuning.

The proposed method presents more insights in the applicability of deep learning methods for detection of phenotype in rare disease. In future work we intend to evaluate a algorithm for automatic detection of a genotype-phenotype correlation.

In conclusion, this study presented for the first time a deep learning model able to identify congenital generalized lipodystrophy phenotype with good accuracy and sensitivity, above 90%, even when using tests with confusing images, with similar characteristics. These deep learning algorithms can be implemented into routine clinical practice, assisting health professionals in the diagnosis of patients with this rare disease. The use of this tool through cell phone applications will facilitate access to this technology, reaching health services in the most remote regions and transforming patient care.

Author contributions

All authors contributed to data analysis, drafting and revising the article, gave final approval of the version to be published, and agree to be accountable for all aspects of the work.

Data availability

The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Conflict of interest

The authors had no conflicts of interest.

Funding

This study did not have funding.

Garg A. Acquired and inherited lipodystrophies. N Engl J Med. 18;350(12):1220-34 (2004).
Nelson, M.D. et al. Cardiac steatosis and left ventricular hypertrophy in patients with generalized lipodystrophy as determined by magnetic resonance spectroscopy and imaging. Am J Cardiol. 1;112(7):1019-24 (2013).
Lima, J.G., Dos Santos, M.C.F., de Melo Campos, J.T.A. Congenital Generalized Lipodystrophy. J Rare Dis Res Treat. 3(2): 1-6 (2018).
Russakovsky, O. et al. ImageNet Large Scale Visual Recognition Challenge. Int J Comput Vis. 115, 211–252 (2015).
Smith, K., Piccinini, F., Balassa, T., Koos, K., Danka, T., Azizpour, H., Horvath, P. Phenotypic Image Analysis Software Tools for Exploring and Understanding Big Image Data from Cell-Based Assays. Cell Syst. 6, 636–653 (2018).
Sánchez Fernández, I. et al. Deep learning in rare disease. Detection of tubers in tuberous sclerosis complex. PLOS ONE. 15(4): e0232376 (2020).
Sernadela, P. et al. Linked Registries: Connecting Rare Diseases Patient Registries through a Semantic Web Layer. BioMed Res. Int. 1–13 (2017).
Ekins, S. Industrializing rare disease therapy discovery and development. Nat. Biotechnol. 35, 117–118 (2017).
About Rare Diseases. Available online: https://www.eurordis.org/about-rare-diseases (accessed on 12 may 2022).
Ronicke, S., Hirsch, M.C., Turk, E., Larionov, K., Tientcheu, D., Wagner, A.D. Can a decision support system accelerate rare disease diagnosis? Evaluating the potential impact of Ada DX in a retrospective study. Orphanet J. Rare Dis. 14, 69 (2019).
Nestler-Parr, S. et al. Challenges in Research and Health Technology Assessment of Rare Disease Technologies: Report of the ISPOR Rare Disease Special Interest Group. Value Health. 21, 493–500 (2018).
Stoller, J.K. The Challenge of Rare Diseases. Chest. 153, 1309–1314 (2018).
Smith, K. et al. Phenotypic Image Analysis Software Tools for Exploring and Understanding Big Image Data from Cell-Based Assays. Cell Syst. 6, 636–653 (2018).
Patni, N., Garg, A. Congenital generalized lipodystrophies--new insights into metabolic dysfunction. Nat. Rev. Endocrinol. v. 11, n. 9, p. 522-34 (2015).
Magre, J. et al. Identification of the gene altered in Berardinelli-Seip congenital lipodystrophy on chromosome 11q13. Nat. Genet. v. 28, n. 4, p. 365-70 (2001).
NOLIS, T. Exploring the pathophysiology behind the more common genetic and acquired lipodystrophies. J. Hum. Genet. v. 59, n. 1, p. 16-23, (2014).
Araújo-Vilar, D. et al. LipoDDx: a mobile application for identification of rare lipodystrophy syndromes. Orphanet J Rare Dis. 15, 81 (2020).
Gurovich, Y. et al. Identifying facial phenotypes of genetic disorders using deep learning. Nat Med. 25, 60–64 (2019).

No competing interests reported.

Identifying congenital generalized lipodystrophy using deep learning- DEEPLIPO

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Materials And Methods

Results

Discussion

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1