A cascaded deep convolution neural network based CADx system for psoriasis lesion segmentation and severity assessment
Introduction
Psoriasis is an irrepressible, and autoimmune skin disease which can be controlled by determined and attentive treatment [1]. Skin cells of psoriasis mature faster than normal and are characterized by reddish lesions covering the silvery-white scales on the skin surface [2]. Although dermatologists predict that generic is the main reason for psoriasis, but the exact reason is yet to be explained [3]. Irrespective of gender, age and racism, it affects almost 125 million individuals worldwide [4]. In Europe, USA, Malaysia, and India, psoriasis affects approximately 6.5% [5], 3.15% [5], 3% [6] and 1.02% [7] of the population respectively. Patients affected with psoriasis undergo a painful and discomfort life, due to the constant itching and ugly physical gesture [8]. The diagnosis of psoriasis is a challenging task due to the high degree of visual similarity between healthy skin and psoriasis lesion. For effective diagnosis of psoriasis, dermatologists aim at identifying the precise segmentation of lesion and its risk taxonomy [1], [2], [3], [4].
In current practice, dermatologist utilizes the subjective measurement which requires visual and haptic cues for psoriasis lesion segmentation and it is severity measurement. However, these methods are time-consuming and might provide unreliable results due to its highly subjective nature [5], [6], [7], [8], [9]. Thus, objective methods have been proposed by various researchers in the recent past. However, the objective methods suffer from limitations related to the low contrast of the skin, uneven illumination condition, similar appearance between the lesion and healthy skin, high intra and inter-class variation of psoriasis skin lesion. Hence, an automated psoriasis risk assessment system can be the eventual solution to segment the lesions and to find the severity level of psoriasis.
In this context, various conventional machine learning techniques have been proposed for psoriasis lesion segmentation and its severity classification in the recent past [10], [11], [12], [13], [14], [15]. However, most of the existing assessment techniques are feature-based and they have validated over a small dataset, which makes them less robust and generalizable [10], [11], [12], [13], [14], [15]. Shrivastava et al. [10], [11], Vani et al. [12], and Murugeswari et al. [13] proposed machine learning-based systems for psoriasis risk assessment, which are entirely dependent on the color and texture features. Lu et al. [14], [15] applied roughness features with decision tree classifier to evaluate psoriasis severity and a combination of Markov random field (MRF) and support vector machine (SVM) to segment the psoriasis lesion. In the reported works on machine learning-based psoriasis diagnosis, the classification performance is heavily dependent on the type of features extracted, which makes them less suitable for clinical studies.
In recent times, contrary to traditional machine learning approaches techniques based on deep learning have been proven to provide models with high reliability for image recognition, segmentation, and classification, since it involves automatic extraction of features invariant to distortion and position of the input images [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27]. In [25], U-Net based fully convolutional neural network has been implemented for the automatic psoriasis skin lesion segmentation of 5241 psoriasis images with the dice coefficient of 93.03%. Jian-Ming et al. [26] have implemented a psoriasis binary classifier based on deep neural network (DNN) using 5700 normal and psoriasis images. In [26], the proposed DNN model only determines whether the input image is psoriasis or not, with the classification accuracy of 91%. None of the deep learning-based schemes have concentrated on all three stages of psoriasis diagnosis i.e. psoriasis recognition, lesion segmentation, and severity assessment.
In this paper, a cascaded deep convolution neural network (DCNN) based CADx system has been proposed which combines all three stages namely; psoriasis recognition, lesion segmentation, and severity assessment in a single framework. For the automatic segmentation of the psoriatic lesions, the modified version of U-Net architecture [25] has been used. Whereas, for recognition and severity assessment, a modified VGG model has been implemented. Among the deep learning architectures, LeNet [28], Alexnet [29], and VGG networks [30] have been successfully used for image classification applications. The use of VGG-16 in the present work is motivated by the fact that it allows a uniform architecture for extracting features from a similar kind of skin lesion-based images such as melanoma detection [27]. However, the standard VGG-16 network has a large number of trainable parameters which increases the computational burden and quite often leads to overfitting. In this regard, standard VGG-16 with 16 weight layers has been modified into 9 weight layer ConvNet configuration.
The major contributions of this paper can be summarized as (i) Design of fully automated deep learning-based CADx system for performing all the intended tasks in psoriasis diagnosis i.e., psoriasis recognition, lesion segmentation, and its severity assessment in a single framework, (ii) The standard VGG-16 model has been modified and successfully implemented for psoriasis recognition and psoriatic lesion severity assessment task. Also, the modified U-Net model has been implemented for the psoriatic lesion segmentation task, (iii) The performance of the proposed framework has been evaluated for each stage of diagnosis on an extensive dataset using k-fold cross-validation procedure, along with benchmarking against previously reported systems (iv) The appropriateness of the psoriasis severity assessment of the proposed CADx system performance was validated and compared using three different experiments: without lesion segmentation, with manual lesion segmentation, and with automatic lesion segmentation.
The paper is organized as follows. Section 2 presents a brief overview of all the stages of the proposed CADx system. The details of the dataset and system implementation are presented in Section 3. The experimental results of each stage of the CADx system are presented in Section 4. Further discussions on the proposed CADx system are carried out in Section 5. Finally, Section 6 concludes the paper.
Section snippets
Overview of proposed CADx system
In this study, we have presented a fully integrated CADx system which includes, three stages namely, recognition, segmentation and severity assessment of psoriasis skin lesion. In the first stage, an automated psoriasis skin lesion recognition algorithm has been proposed which utilizes an efficient CNN based architecture. In the second stage, a segmentation module based on a modified fully convolutional network architecture has been employed to precisely segment the recognized psoriatic lesion.
Dataset
The psoriasis image dataset was collected from both male and female psoriasis patients irrespective of age and community. Images were captured by using a Sony NEX-5 with a 22 mm lens and 350 dpi at Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India in the presence of dermatologist. The images were then processed in the Joint Photographic Expert Group (JPEG) format with a color depth of 24 bits per pixel. A total of 5000 healthy skin images and 5000 unhealthy (having
Evaluation metrics
The performance of the individual stages of the proposed CADx system has been evaluated using various metrics such are accuracy, sensitivity, specificity, precision, F1-score, Jaccard index, dice coefficient, and AUC. Accurate detection of the lesion, as well as the severity classification with respect to the ground truth, has been quantified by using accuracy which is calculated by using Eq. (7). where; TP true positive, TN true negative, FP false positive and FN
Comparison with existing studies for psoriasis severity assessment
Various techniques have been proposed for psoriasis severity assessment with and without automatic lesion segmentation in the recent past. However, to the best of our knowledge, no work has been reported on the development of CADx system using deep learning for fully automated psoriasis lesion segmentation and severity assessment in a single framework with validation on a large dataset. The proposed system has been compared with existing works [11], [13], [15], [26] in Table 6. The existing
Conclusion
In this paper, we proposed a cascaded DCNN based automated CADx system for psoriasis recognition, segmentation, and severity assessment. In the proposed CADx system, all the three stages are seamlessly integrated into a single framework and the whole intervention is performed in an automated way without any manual interactions. The adopted U-Net for automated segmentation has two advantages (i) it made the system more reliable and automatic; (ii) it also helped to increase the severity
CRediT authorship contribution statement
Manoranjan Dash: Designed and performed experiments, analyzed data and co-wrote the paper. Narendra D. Londhe: Conception, design of the study and supervised the research. Subhojit Ghosh: Supervised the research and co-wrote the paper. Ritesh Raj: Designed experiments and revised the manuscript thoroughly. Rajendra S. Sonawane: Acquisition of data, analysis and interpretation of data.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors would like to thank the psoriasis patients who participated in this study and Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India , for their support.
References (38)
- et al.
The annual cost of psoriasis
J. Am. Acad. Dermatol.
(1993) - et al.
Geoepidemiology and environmental factors of psoriasis and psoriatic arthritis
J. Autoimmun.
(2010) - et al.
Computer-aided measurement of psoriatic lesion area in a multicenter clinical trial comparison to physician’s estimations
J. Dermatol. Sci.
(2006) - et al.
A novel approach to multiclass psoriasis disease risk stratification: Machine learning paradigm
Biomed. Signal Process. Control
(2016) - et al.
A novel and robust Bayesian approach for segmentation of psoriasis lesions and its risk stratification
Comput. Methods Programs Biomed.
(2017) - et al.
PsLSNet: Automated psoriasis skin lesion segmentation using modified U-Net-based fully convolutional network
Biomed. Signal Process. Control
(2019) Handbook of Psoriasis
(2008)Genetics of psoriasis
Arch. Dermatol. Res.
(1998)- et al.
Psoriasis in India: Prevalence and pattern
Indian J. Dermatol. Venereol. Leprol.
(2010) - et al.
The impact of psoriasis on quality of life: results of a 1998 National Psoriasis Foundation patient-membership survey
Arch. Dermatol.
(2001)
The risk of depression anxiety and suicidality in patients with psoriasis: a population-based cohort study
Arch. Dermatol.
Evaluating new therapies for psoriasis
Manage. Care
Classification of skin lesions based on texture features
Int. J. Eng. Res. Appl.
Fuzzy based visual texture feature for psoriasis image analysis
Int. J. Comput. Inf. Eng.
Automatic scoring of erythema and scaling severity in psoriasis diagnosis
Automatic segmentation of scaling in 2-d psoriasis skin images
IEEE Trans. Med. Imaging
Deep Learning
Auto-context convolutional neuralnetwork (auto-net) for brain extraction in magnetic resonance imaging
IEEE Trans. Med. Imaging
Automatic skin lesion segmentation using deep fully convolutional networks with jaccard distance
IEEE Trans. Med. Imaging
Cited by (17)
Automatic SCOring of Atopic Dermatitis Using Deep Learning: A Pilot Study
2022, JID InnovationsClassification of cervical cells leveraging simultaneous super-resolution and ordinal regression
2022, Applied Soft ComputingCitation Excerpt :By optimizing features from the segmented cytoplasm and nucleus, high classification accuracies (e.g., 96.8%) were achieved [28,29] while these high values would degrade once the automated segmentation error was taken into account [30,31]. Recently, representation learning has been popularized by CNN methods [32–34] which have also significantly improved the performance in a variety of medical imaging applications [35–42] such as classification of cervical cells in cytology images [43]. Different CNN models were applied to classify and identify cervical cells in [44–46] . [1],
A convolutional neural network trained with dermoscopic images of psoriasis performed on par with 230 dermatologists
2021, Computers in Biology and MedicineCitation Excerpt :However, current studies about the computer-aided diagnosis of skin diseases focus mainly on the classification of pigmented skin tumors. Only a few researchers have carried out artificial intelligence (AI) research on the identification and severity assessment of psoriasis based on clinical images [12,13]. Therefore, we aim to train an efficient deep-learning network to recognize dermoscopic images and achieve a fine-grained classification covering a variety of papulosquamous diseases.
Automated psoriasis lesion segmentation from unconstrained environment using residual U-Net with transfer learning
2021, Computer Methods and Programs in BiomedicineCitation Excerpt :The deep learning models require a deeper network to extract important features from complex images [49–55]. But, sometimes the training of a deeper network with a large number of convolutional layers suffers from a vanishing gradient problem, which degrades the model performance [56–58]. This problem is solved by using a residual network with a short skip connection [59], which helps in the effective training of a deeper model.
Objective scoring of psoriasis area and severity index in 2D RGB images using deep learning
2024, Multimedia Tools and Applications