Developed Newton-Raphson based deep features selection framework for skin lesion recognition
Introduction
In skin cancer, melanoma is the deadliest type – responsible for the death of large number of people worldwide [1], [2]. Skin cancer can be treatable if diagnosed at the early stages, otherwise, consequences will be severe [3]. At the early stage, melanoma starts in the melanocyte cells, which seems like a mole having black or brown color [4], [5]. In the year 2017, reported skin cancer cases, in United States (US) only, are 95,360 (57,140 men and 38,220 women), in which melanoma cases are 87,110 (52,170 men and 34,940 women). The estimated deaths occurred in the USA since 2017 are 13,590 (9250 men and 4340 women) [6], [7]. In the year 2018, an estimated 99,550 (60,350 men and 39,200 women) cases are reported. From those melanoma cases are 91,270 including 55,150 men and 36,120 women. The death cases, on the other hand, in 2018 are 13,460 including 9070 men and 4390 women.
In 2019 only, stats show 104,350 cases, including 62,320 men and 42, 030 women. The number of melanoma cases are 96,480, including 57,220 men and 39,260 women. The number of melanoma death cases during 2019 in USA are 7320 [8]. Fig. 1 provides cancer statistics in brief for the year 2015 to 2019.
A conventional method for skin lesion detection is through visual inspection, which is quite a challenging task – clearly depends upon the expert. A dermatologist mostly uses the screening methods such as 7-point checklist [9], ABCDE rule [10], and several other advanced techniques like optical imaging system and light, etc. [11] for the detection of skin lesion. These methods perform well but are time-consuming and are also not free from human error. Due to the recent advancements in the field of computer vision (CV), several computerized systems are utilized in clinics which play helps to doctors for diagnosis at early stage. Most of the existing methods incorporates four primary stages for skin lesion detection from contrast stretching to classification [12], [13].
Preprocessing step is very important for the removal of noises such as hair, bubbles, etc., and also plays a vital role in an accurate segmentation [14]. Feature extraction from the segmented image is a crucial step, as good features lead to an accurate classification and vice versa [15], [16]. Lately, with the advent of deep learning methods [17], [18], [19], there is an increasing trend to utilize them in medical domain [20]. By embedding the concept of transfer learning, convolutional neural network (CNN) models [21], [22] - trained on the large image datasets are retrained on the skin datasets. Feature selection is an important research in the area of machine learning and CV [23], [24]. In the medical imaging, the extraction of features from raw images generates various patterns information and few of them are not essential for classification task [25]. The irrelevant information misguides the selected classifiers and reduces the overall performance.
Inspired from the comparative work by Fernandes et al. [26], in which early skin lesion is detected based on two state-of-the-art techniques, color constancy, and skin lesion analysis. Authors performed a detailed analysis to conclude that color constancy approach is a better choice for skin lesion detection. Additionally, they also concluded that early detection of skin lesion is quite expedient for the treatment of melanoma.
In this work, implemented a DenseNet pre-trained CNN model [27] for deep feature extraction and later best most discriminant features are selected by employing a Newton Raphson (NR) method. Our major contributions are- (a) Artificial Bee Colony (ABC) based an efficient contrast stretching method is proposed for an accurate segmentation; (b) Faster RCNN is implemented for lesion detection – utilizing ground truth pixels’ information; (c) An entropy based activation function for deep features extraction is implemented, and (d) A Newton Raphson (IcNR) computational method is implemented for the most discriminant features selection.
The remaining manuscript is ordered as follows: Related work is described in Section 2. Proposed DLNR method presented in Section 3. Results and comparison are discussed in Section 4. Finally, Section 5 concludes the overall manuscript.
Section snippets
Related Work
An automatic mechanism for the recognition of skin lesion is an arduous task due to a set of factors including low contrast, irregularity, presence of several artifacts like hairs and bubbles, etc. Manual inspection of a skin lesion is dependent on a qualified specialist, which can't be available whole time, therefore, machine learning based methods are proposed by a pool of researchers working in this domain. Codella et al. [28] presented a hybrid method for lesion recognition, ensemble deep
Proposed methodology
The proposed deep learning and NR (DLNR) based skin lesion recognition system consists of the following primary steps, Fig. 2, where contrast stretching is performed to visually improve the lesion area. Later, Faster RCNN (F-RCNN) is applied on contrast stretched images for lesion boundary localization. After localization of lesion boundary, extract the deep features by employing pre-trained CNN model name DenseNet. Transfer learning based optimized skin lesion features is extracted that later
Datasets description
The proposed skin lesion recognition method is tested on two datasets, ISBI 2016 [46] and ISBI 2017 [47]. The ISBI 2017 dermoscopy dataset includes total of 2750 RGB images of different resolutions. From total dermoscopy images, 517 images are malignant and 2223 are benign. The ISBI 2016 dermoscopy dataset includes total of 1279 images of different resolutions having 273 malignant and 1006 benign.
Simulation procedure
In the simulation procedure, images are divided into training and testing images. Two different
Conclusion
A new automated system is proposed for skin lesion localization and recognition – utilizing the concept of deep learning and IcNR based feature selection. The proposed IcNr selection method is evaluated on two freely available datasets- ISBI2016 and ISBI2017 to achieving an average accuracy of 94.5% and 93.4%, respectively. From the results, it is concluded that the contrast stretching step increases the segmentation accuracy by enhancing the lesion area compared to the background.
Declaration of Competing Interest
On the behalf of corresponding author, all authors declare that they have no conflict of interest.
References (56)
- et al.
CCDF: Automatic system for segmentation and recognition of fruit crops diseases based on correlation coefficient and deep CNN features
Comput. Electron. Agric.
(2018) - et al.
Appearance based pedestrians’ gender recognition by employing stacked auto encoders in deep learning
Future Gener. Comput. Syst.
(2018) - et al.
Skin lesion analysis towards melanoma detection using deep learning network
Sensors
(2018) - et al.
Image contrast enhancement using an artificial bee colony algorithm
Swarm and Evolutionary Computation
(2018) - et al.
Big data analysis for brain tumor detection: Deep convolutional neural networks
Future Gener. Comput. Syst.
(2018) - et al.
A method for the detection and classification of diabetic retinopathy using structural predictors of bright lesions
J. Comput. Sci.
(2017) - et al.
An implementation of normal distribution based segmentation and entropy controlled features selection for skin lesion detection and classification
BMC Cancer
(2018) - et al.
Skin lesion segmentation and recognition using multichannel saliency estimation and M-SVM on selected serially fused features
J. Ambient Intell. Human. Comput.
(2018) - M.A. Khan, T. Akram, M. Sharif, K. Javed, M. Rashid, and S.A.C. Bukhari, "An integrated framework of skin lesion...
- et al.
Melanocytic and nevus lesion detection from diseased dermoscopic images using fuzzy and wavelet techniques
Soft Comput.
(2018)