Gabor wavelet-based deep learning for skin lesion classification
Introduction
Cancer is one of the main problems worldwide [1]. The disease can take in several forms and occurs in different locations of the human body. One of the most commonly seen and deadly cancer types in women is breast cancer. Another widely recognized and cancer type with high mortality rates among men is prostate cancer. Melanoma skin cancer is also common and deadly among men and women. This type of cancer is known as the most commonly treated skin cancer type, and of the population is affected in the United States. Moreover, it has also been reported that melanoma skin cancer type is the leading cause of cancer-related deaths in the United States. A recent study [1] has reported the estimated number of new cases and new cancer-related deaths.
Melanoma is one of the skin cancer types, and its associated mortality rate is higher than other types of skin cancer. This malignant type also causes more deaths than different skin cancer types, [[2], [3], [4]]. Melanoma and nevus are known as melanocytic skin lesions. Non-melanocytic skin lesions also occur in malignant and benign forms. Basal cell and squamous cell carcinoma are non-melanocytic malignant skin lesions. Actinic keratosis is the early form of squamous cell carcinoma. Dermatofibroma, vascular, and benign keratosis are known as non-melanocytic benign lesions. Although Basal-cell and squamous-cell carcinoma are more common, melanoma causes a greater number of deaths, [2,5].
Early detection of skin cancer leads to effective treatment. Early detection of cancer could be achieved via automated detection of cancer types using deep learning techniques. Recent studies show that accurate automatic detection of skin cancer can be achieved using deep convolutional neural networks (CNNs). Recent surveys [[6], [7], [8]] have also explained the automated detection of skin cancer forms.
Esteva et al. [1]. used a single GoogleNet CNN model [9,10] for skin lesion prediction. Han et al. [2]. also utilized the ResNet-152 CNN model [11] for skin cancer type recognition. Kaymak and Serener [12] used several single CNN models for the prediction of different categories of skin lesions. On the other hand, several studies have used an ensemble of CNN models for skin cancer detection. Hagari [13] combined AlexNet [14], GoogleNet, VGG [15], RestNet for lesion prediction. Mahbod [16] proposed to fusion of AlexNet, ResNet-18, ResNet-100 and VGG16 for malignant type lesion classification.
The proposed method is different than the above mentioned methods. The proposed new model contains several Gabor wavelet and skin image-based CNN models. This method is an extension of previous techniques and utilizes Gabor representations in a unified CNN model. The model also allows the integration of image data for further enhancements. The Gabor wavelet-based deep convolutional neural network method utilizes Gabor filters on skin images to obtain detailed representations of the skin lesions. Then these details are modeled using deep convolutional neural network models. Furthermore, the fusion of skin images and Gabor representations is performed to increase the accuracy of the proposed method. All details are modeled using deep networks, and then outputs are fused using a probabilistic fusion method to detect skin cancer types.
Previous studies [12,13,16,21] have used skin images as inputs and an ensemble of CNN models for the prediction of skin lesion types. In contrast, this work is based on the decomposition of input images into seven directional sub-bands and then seven sub-band images along with the input image are used as inputs to eight parallel CNNs channels. Eight probabilistic predictions originating from the input channel and seven directional sub band channels are fused for the prediction of skin lesion type.
The main novelties of this work can be summarized as follows. Primarily, a novel Gabor wavelet-based deep convolutional neural network is proposed for the detection of malignant melanoma and seborrheic keratosis. The proposed method is based on the decomposition of input images into seven directional sub-bands. Gabor based approach provides directional decomposition where each sub-band gives isolated decisions that can be fused for improved overall performance. Seven sub band images and the input image are used as inputs to eight parallel CNNs to generate eight probabilistic predictions. Decision fusion based on the sum rule is utilized to classify the skin lesion. Gabor based approach provides directional decomposition where each sub-band generates isolated directional information in the respective sub-band channel. CNN processing in each channel generates learning in the respective direction which can be regarded as decorrelated from the other directions. Hence, combining the decisions of these channels improve the overall classification performance. This approach can be regarded as an augmentation method coupled with multi-channel CNN learning.
The organization of this paper is as follows. The related work section describes previous studies on this subject. In Section 3, we describe the proposed method and give explanations for both the single and ensemble of deep convolutional neural network models. This section also explains Gabor representations and model generations. In Section 4, we report the proposed model evaluation. This section also compares the results of the proposed and other methods. Finally, we conclude the proposed methodology in Section 5.
Section snippets
Related work
There are two different classification approaches for skin lesion classification. These categories are single deep learning-based methods and the ensemble of deep learning methods.
Method
The proposed method builds on Gabor wavelet-based CNN models and an image-based CNN model, Fig. 1. First, seven directional Gabor wavelets represent skin images. Then, the seven Gabor-based CNN models are learned to recognize skin lesion classes. The image-based CNN model is also generated using only skin images. Finally, the CNN model outputs are combined for the estimation of melanoma and seborrheic keratosis classes (see Fig. 2).
Decision fusion allows a combination of output probabilities of
Performance evaluation
The performance of the proposed Gabor wavelet-based deep learning model is evaluated for melanoma and seborrheic keratosis recognition tasks. The area under the receiver operating characteristic curve (M-AUC), accuracy (M-ACC), sensitivity (M-SE), and specificity (M-SP) shows the performance of the melanoma. Similarly, the area under the receiver operating characteristic curve (SK-AUC), accuracy (SK-ACC), sensitivity (SK-SE), and specificity (SK-SP) defines the performance of seborrheic
Conclusion
This work proposes a novel Gabor wavelet-based deep learning model for melanoma and seborrheic keratosis. This model builds on an ensemble of seven Gabor wavelet-based CNN models. Furthermore, this model fuses the Gabor wavelet-based model and an image-based CNN model. The performance evaluation results show that an ensemble of the image and Gabor wavelet-based model outperforms a single image and Gabor wavelet-based models. This ensemble also outperformed the group of only Gabor wavelet-based
References (29)
- et al.(2018)
Skin lesion classification with ensembles of deep convolutional neural networks
Biomed. Inf.
(2018)- et al.
Fusing fine-tuned deep features for skin lesion classification
Comput. Med. Imag. Graph.
(2019) - et al.(2018)
- et al.
Cancer statistics
Cancer J.
(2018) - et al.
Dermatologist-level classification of skin cancer with deep neural networks
Nature
(2017) - et al.
Classification of the clinical images for benign and malignant cutaneous tumors using a deep learning algorithm
J. Investig. Dermatol.
(2018) - et al.
The HAM10000 Dataset: A Large Collection of Multi-Source Dermatoscopic Images of Common Pigmented Skin Lesions
(2018) - et al.
Update on keratinocyte carcinomas
N. Engl. J. Med.
(2018) - et al.(2015)
Dermoscopy image analysis: overview and future directions
IEEE J. Biomed. Health Inf.
Going deeper with convolutions
Rethinking the inception architecture for computer vision
Resnet in Resnet: Generalizing Residual Architectures
Cited by (78)
Fractional WSD: Fractional war strategy dingo optimization with unified segmentation for detection of skin cancer
2024, Biomedical Signal Processing and ControlThe power of deep learning for intelligent tumor classification systems: A review
2023, Computers and Electrical EngineeringA survey, review, and future trends of skin lesion segmentation and classification
2023, Computers in Biology and MedicineAn adaptive weight search method based on the Grey wolf optimizer algorithm for skin lesion ensemble classification
2024, International Journal of Imaging Systems and TechnologyEffective melanoma classification using inter neighbour mean order interleaved pattern on dermoscopy images
2024, Multimedia Tools and ApplicationsFine-tuning pre-trained neural networks for medical image classification in small clinical datasets
2024, Multimedia Tools and Applications