CNN Models for Skin Lesions Detection

CNN Models for Skin Lesions Detection: History

View Latest Version

Please note this is an old version of this entry, which may differ significantly from the current revision.

Subjects: Computer Science, Artificial Intelligence

Contributor:

Muhammad Azeem

Kaveh Kiani

Taha Mansouri

Nathan Topping

Skin cancer is a widespread disease that typically develops on the skin due to frequent exposure to sunlight. Although cancer can appear on any part of the human body, skin cancer accounts for a significant proportion of all new cancer diagnoses worldwide. There are substantial obstacles to the precise diagnosis and classification of skin lesions because of morphological variety and indistinguishable characteristics across skin malignancies. Deep learning models have been used in the field of image-based skin-lesion diagnosis and have demonstrated diagnostic efficiency on par with that of dermatologists.

deep learning
convolutional neural network
computer vision
computer-aided diagnosis
skin lesion
skin cancer
melanoma
medical imaging

1. Introduction

“Cancer”—the collective term by which a group of linked diseases is referred to—occurs when several body cells start to divide uncontrollably and to invade nearby tissues [1]. Skin cancer is one of the most common forms of cancer [2] and commonly occurs when the skin is frequently exposed to sunlight [3]. Ultraviolet rays, which are the primary cause of skin cancer, harm the DNA in skin cells [4]. There are three main types of skin cancer: basal cell carcinoma, squamous cell carcinoma, and melanoma [5]. However, non-melanoma skin cancers present a lower risk of spreading to other parts of the body and are easier to treat than melanoma [6]. It is estimated that while melanoma only accounts for 4% of skin cancer cases, it causes 75% of skin-cancer-related deaths [7]. Globally, there were an estimated 287,700 new cases of melanoma in 2018 and an estimated 60,700 deaths from melanoma in the same year [2]. The incidence of melanoma has grown significantly in recent times, partly due to an increase in sun-seeking behaviors [7]. For example, in the United States, the lifetime risk of developing malignant melanoma has increased from 1 in 5000 in the year 1935 to 1 in 74 in the year 2000 [8].

Early-identified individuals have a better chance of recovering, because the five-year survival rate for patients with early-identified malignant melanoma is 94% [8]. Early diagnosis is therefore a critical factor in reducing skin cancer mortality. Dermoscopy is a specialist technology that produces high-resolution magnified images of the skin by controlling light and removing surface skin reflectance [9], and the clinical use of dermatoscopic images has the potential to improve diagnosis rates for melanoma [8] and, ultimately, to save lives [10]. However, given the existing pressures on healthcare systems, cost-effective strategies are needed to facilitate increased screening and diagnosis [11], and there has been significant interest in computer-aided diagnosis [6]. An active area of research is the use of smartphone applications incorporating machine-learning methods to analyze images and assist in early melanoma diagnosis, whilst reducing pressures on healthcare systems and clinical staff [12].

Clinicians have traditionally utilized the ABCD guidelines to differentiate melanoma from non-malignant skin lesions, with A denoting asymmetry, B denoting border irregularity, C denoting color variations, and D denoting diameter greater than 6 mm [8,13]. However, due to the morphological variation and complex characteristics of skin lesions, there can be challenges with inter-observer and intra-observer concordance, further motivating the exploration of computer-aided diagnosis techniques [14]. By utilizing texture cues, geometrical aspects, color features, and combinations of these features, medical images can be used to identify and classify skin cancer conditions [15]. The use of traditional machine-learning techniques in melanoma diagnosis has typically involved feature extraction from dermatoscopic images, to build a set of relevant features that can be used to train a classification model, often utilizing the ABCD rule to define an appropriate feature set [14,16,17,18]. Due to the complexity of skin lesions, it is difficult for researchers to recognize skin malignancies by using these geometrical properties [18].

In the field of image-based melanoma diagnosis, the development of deep learning and, in particular, convolutional neural networks (CNNs) has reduced reliance on manual-feature extraction techniques. CNN-based classification methods have also demonstrated diagnostic effectiveness comparable to that of dermatologists [19]. In [20], researchers mainly concentrated on the automatic identification and categorization of skin cancer, as computer-aided screening technologies had become more prevalent. Many studies have been conducted to categorize melanoma skin lesions, using CNNs on various datasets, including MNIST HAM10000 [21], the International Skin Imaging Collaboration 2018 (ISIC) [22], the PH2 public database [23], and the International Symposium on Biomedical Imaging (ISBI) [24]. Using these datasets, promising results have been achieved, employing a variety of pre-trained CNN models, including ResNet50 [25], ImageNet50 [26], and DenseNet201 [27] alongside additional cutting-edge models. Some datasets, such as the Dermatological and Surgical Assistance Program at the Federal University of Espirito Santo (PAD-UFES-20) dataset [28], have not been explored as extensively for the identification of skin lesions.

2. CNN Models for Skin Lesions Detection

Machine learning and AI have made significant advances in cancer prediction and detection in recent years [29]. Dermoscopy is a non-invasive imaging method for taking comprehensive images of skin lesions [30]. The development of computer-aided diagnosis systems has been spurred by research into the need for accurate and early identification of skin illnesses, including melanoma and other types of skin cancer [31]. When used to automate the classification of skin lesions based on dermoscopy images, deep learning methods, in particular CNNs, have demonstrated encouraging results [32]. Earlier methods for classifying skin lesions relied on manually engineered characteristics and conventional machine-learning techniques [33]. The use of deep learning techniques, particularly CNNs, has reduced the reliance on manual feature extraction in skin-lesion classification [34]. For classification tasks involving skin lesions, well-known CNN architectures like AlexNet [35], VGGNet [36], and InceptionNet [37] have been adapted and refined.

Deep learning models were developed by [38], where CNN models were trained and evaluated on the HAM10000 dataset, which delivered 90 per cent validation accuracy when classifying various forms of skin malignancies. In [39], with the help of a data augmentation technique, a CNN classification model was proposed and trained, using a public dataset of skin lesions that included 600 test and 6162 training images, achieving a classification accuracy of 89.2%. Using a cutting-edge prediction algorithm, benign and malignant skin lesions were separated into categories in [40] with a CNN and a novel regularizer. The model was then trained with a dataset obtained by the International Skin Imaging Collaboration (ISIC) databank, which acquired a summed accuracy score of 97.49%. In [41], by using fuzzy C-means clustering and K-means clustering, the researchers classified skin lesions using a CNN model trained with the ISIC dataset, achieving an accuracy of 98.83%.

In [42], transfer-learning techniques were used with two CNN architectures, ResNet50 and DenseNet169. The models were trained and validated on the HAM10000 dataset, and the highest-performing generated an accuracy score of 91.2%. In addition to existing methods, such as border extraction utilizing XOR with regression logic, another CNN model was suggested in [43]. The datasets from PH2 and ISBI 2017 were utilized to train the model, which achieved a 97.8% accuracy rate. In [44], to enhance the performance of the proposed CNN model, a transfer-learning strategy was employed, using a publicly available dataset on Kaggle, which resulted in accuracy of 79.45%. Another CNN model was designed by [45] and trained with a medical dataset acquired from Al-Kindi Hospital and Baghdad Medical City to classify skin lesions, obtaining accuracy of 89%. In [46], seven different types of skin problems were categorized, using a CNN.

In [47], a U-Net-based model was proposed for semantic segmentation of skin-lesion images, and the proposed model was validated against ISIC2018, ISIC2017, and PH2. In [48], a U-Net model was also used for skin lesion semantic segmentation and was evaluated with ISIC2017 and ISIC2018, with accuracy of 94.9% and 95.4%, respectively. In [49], a CNN model that distinguished blemishes into moderate skin cancer and cases of acne was developed, using images of diverse benign skin cancers and acne cases, and it yielded precision of 96.4%. In [50], a dataset of skin cancer dermoscopy images was utilized, subjected to a number of data-cleaning steps to reduce noise and enhance the quality of images, and then a CNN model was employed for categorization, achieving accuracy of 98.38%. In [51], the HAM10000 dataset was used to train a Siamese neural network. While the classification accuracy was lower than with some other models, this approach was able to detect examples that did not belong to the training classes.

In [52], the PH2 dataset of dermoscopic images was utilized to create and build a CNN model, which achieved test-set accuracy of over 95%. In [53], a six-layer CNN model was created and trained on the ISIC dataset and showed promise, with accuracy of 89.30% in classifying skin lesions. Another State-of-the-Art CNN model was designed and developed by [54]. The model achieved 97.50% accuracy results when used with the ISIC and PH2 datasets, to separate skin lesions.

In [55], along with data augmentation and image preparation procedures, a CNN model, which obtained 95.2% accuracy, was trained and tested on the HAM10000 dataset. In [56], a CNN model was designed and trained on the ISIC2019 dataset, successfully classifying eight types of skin malignancies with a 94.92% test-accuracy score. In [57], the DenseNet201 model was fine-tuned and trained on the HAM10000 dataset, to classify skin lesions in dermoscopy images, obtaining 86.91% test accuracy. In [58], a deep CNN was created and trained on the ISBI 2017 dataset, to classify melanoma skin lesions. This network achieved 87% accuracy on test data.

Table 1 presents detailed performance metrics and a comparative analysis of the implemented CNN models on each dermoscopy dataset.

Table 1. Comparison of implemented CNN models for skin lesions detection on different dermoscopy datasets.

Ref.	Model	Dataset	Accuracy	Comments
[38]	Deep convolutional neural network	HAM10000	90%	One of the main benchmark datasets was used in this paper, which produced promising results while using a CNN model. However, hyperparameters tuning is required, to increase the accuracy results.
[59]	Convolutional neural network (CNN)	International Skin Imaging Collaboration (ISIC)	97.49%	A CNN model was used, which showed relatively good results, but the size of the dataset needs to be maximized.
[42]	ResNet50	MNIST: HAM10000	91%	A State-of-the-Art model was used, which produced reasonable results on the given dataset. However, the dataset needs to be preprocessed well before training, to obtain more accurate and promising results.
[43]	Deep convolutional neural network	International Symposium on Biomedical Imaging (ISBI)	97.8%	A CNN model was trained on an internationally recognized benchmark dataset. However, the size of the dataset was decreased, which showed good results but could lead to model overfitting.
[48]	U-Net	International Skin Imaging Collaboration (ISIC)	94.9%	A State-of-the-Art model was used, which showed promising results, but more data preprocessing or augmentation is needed for accurate prediction.

This entry is adapted from the peer-reviewed paper 10.3390/cancers16010108

© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license; additional terms may apply. By using this site, you agree to the Terms and Conditions and Privacy Policy.