Keywords

1 Introduction

Skin cancer is the leading form of cancer in the United States [1]. It forms when skin cells multiply abnormally and can prove fatal if it is allowed to metastasize to other areas of the body through the lymphatic system. Most skin cancers result from exposure to Ultraviolet (UV) light. When the skin is unprotected, UV radiation damages DNA and can produce genetic mutations, which can subsequently lead to cancerous growths [2]. According to Didona et al. [3] the most common types of skin cancer in Caucasian populations are melanoma and nonmelanoma (i.e., basal and squamous cell carcinoma) skin cancers (NMSC), with melanoma accounting for 4% of all deaths from cancer [4].

Two methods are commonly employed to diagnose whether a skin sample (biopsy) should be taken: Visual examination of the skin by a physician [5, 6]; or dermatoscopy [7] and/or epiluminescence microscopy by a trained clinician [8]. Thus, initial diagnostic efficiency currently depends exclusively on the competence and perceptual capabilities of the practitioner. Perhaps unsurprisingly, both methods have been found to result in suboptimal detection efficacy [9] with false positives abounding. Hence there exists an urgent need for a screening method with increased sensitivity and specificity to be developed.

To address these issues, medical practitioners have increasingly been seeking to employ automated image processing tools that can more effectively diagnose skin cancer [10]. Maier et al. [11] successfully used dermatoscopic images to train an Artificial Neural Network to differentiate deadly melanomas from melanocytic nevi. Although promising, this study, like earlier attempts [12], was hampered by small sample sizes and a lack image variation [13].

Recent increases in data availability, paired with technological advances, have revigorated these efforts. A deep learning approach was successfully employed and returned more accurate diagnoses than most trained experts [14, 15]. Gautman et al. [16] issued an automation challenge to modelers and reported the top submission had an accuracy of 85.5% for disease classification. More recently, a deep convolutional neural network (CNN) model known as MobileNetV2 using a transfer learning method classified benign versus malignant lesions with an accuracy of 91.33% [17].

The objective of the current research is to expand earlier efforts in developing automated skin cancer detection systems by producing a model capable of accurately classifying seven different types of skin lesions. Stakeholders in this endeavor are patients with skin lesions and practitioners. Prefacing our findings, we demonstrate that our approach can provide a high degree of accuracy (95%) in the early diagnosis of skin cancer(s). Importantly, because human perception is not required, we argue this approach should greatly minimize the negative impact of human factors.

2 Dataset

The dataset was compiled by the Medical University of Vienna [18] and includes 10,015 images of pigmented skin lesions. Images were sampled equally from male and female patients with an average age of 51. Images were collected from different parts of the body (e.g., face, ear, and neck) and captured in resolutions ranging from 8 × 8 pixels to 450 × 600 pixels. Figure 1 displays a sample of the images used in the study. Images fall into seven different classifications:

  • Melanoma (mel): The most dangerous form of skin cancer which generally develops from pigment-containing cells known as melanocytes [19].

  • Basal cell carcinoma (bcc): This cancer affects the basal cells which are responsible for the production of new skins. While it rarely metastasizes it does spread easily [20].

  • Actinic keratosis (akiec): This “pre-cancer” indicator appears as a scaly patch resulting from accumulated UV exposure [21].

  • Benign keratosis-like lesions (bkl): A benign, painless skin disorder which is mostly associated with aging and exposure to UV light [22].

  • Vascular lesions (vasc): Common birthmarks that can be flat or raised [23].

  • Dermatofibroma (df): Superficial benign fibrous histiocytoma which primarily occur in women [24].

  • Melanocytic nevi (nv): Benign birthmarks and moles that resemble melanoma [25].

Fig. 1
figure 1

A lower extremity sample for a 50-year-old male diagnosed with melanocytic nevei (upper left). A lesion sample from a 60-year-old male diagnosed with melanoma (upper right). A face lesion sample from a 70-year-old female diagnosed with basal cell carcinoma (lower left) and a lesion sample from a 50-year-old female diagnosed with benign keratosis-like lesions (upper right)

Full size image

Importantly, these classes are not unique. Thus, some patients may present with more than one type of lesion. More than 50% of lesion images were confirmed by pathology, while the ground truth for the rest of were either follow-up, expert consensus, or confirmation by in vivo confocal microscopy.

3 Methodology

CNNs are the state of the art in deep learning for image classification [26], and there are numerous applications for CNN medical image analysis [27].

3.1 Image Preprocessing

Images were preprocessed using normalization techniques, for example, scaling image intensity to the range of [0, 1]. To increase processing speed each image was down sampled to 50 × 50 × 3 pixels. Images were unevenly distributed across classes. Thus, to remove potential bias subsets were created by randomly sampling evenly from the seven categories that ensured the complete population was considered. One aspect of preparation of the images was to be assure that no repeated image should be appeared in training dataset, to address this issue, a chi-square distance measurement technique was used [28].

Data augmentation techniques were employed to increase the number of images available to train on. Images were rotated, zoomed, and flipped.

3.2 CNN

The architecture of CNN model employed is shown in Fig. 2. It consists of two convolutional parts: First, two convolutional layers followed by a pooling layer with a dropout rate of 0.25; second, two convolutional layers, a pooling layer and dropout rate of 0.30, trailed by a flattening of densely connected layers. The convolutional, also known as pooling steps, condense information. The lowest resolution images did not provide enough information to allow for the second convolutional/pooling layers and were omitted. Sometimes results from the first convolutional models were good, besting more complex models. This pattern was particularly true for medium-resolution images. It appears that medium-resolution images can run out of the information required by more complex models, and performance begins to suffer.

Fig. 2
figure 2

Visualization of the CNN model built

Full size image

3.3 VGG-Net

The last set of models considered employed the VGG16 algorithm [29]. The general structure of this network is a 16-layer CNN that uses 3 × 3 filters with stride and pad of 1, along with 2 × 2 max pooling layers with stride of 2. The convolutional layers have 16, 64, 128, 256, 512 nodes successively. As the spatial size of the input volumes at each layer decrease, the result of the convolutional and pool layers, the depth of the volumes increases as the number of filters increases, doubling after each maxpool layer. The flattened layers consist of 1098, 4098, and 7 nodes. The final layer employs a SoftMax activation function.

4 Results

In the current analysis the CNN was found to produce an accuracy of 93% and minimal test loss of 0.18%. However, it did not sufficiently address the issue of overfitting. This can be seen clearly in the wide gaps between training and test set performance in Fig. 3.

Fig. 3
figure 3

Accuracy (left panel) and loss (right panel) of the CNN model without data augmentation

Full size image

When compared to the CNN without data augmentation, the model including augmented data did improve accuracy to 94% and decreased loss to 0.14% (shown in Fig. 4). The problem of overfitting was not ideal.

Fig. 4
figure 4

Accuracy (left panel) and loss (right panel) of the CNN model with data augmentation

Full size image

The VGG16 had an average accuracy of 93.67%, sensitivity of 95.66%, and specificity of 80.43%. A ten-fold cross validation was used to estimate the efficiency of the model. The learning curve of this topology is shown in Fig. 5. The learning curves indicate that the training loss decreases to a point of stability, and the small gap with training loss suggest overfitting was mostly resolved.

Fig. 5
figure 5

Accuracy (left panel) and loss (right panel) of the VGG16 model

Full size image

Metrics for each k-fold of the model are shown in Table 1. From the table, the ability of the model to correctly identify those with cancer (i.e., true positive rate) is as high as 96% and never lower than 94%, while the ability to correctly identify those without cancer (i.e., true negative rate) is as high as 83% and no lower than 70%. Notably higher than those provided by experts [9].

Table 1 K-fold cross validation metrics
Full size table

Table 2 indicates how accurately each of the seven classes of skin lesions are predicted. The most common type of skin cancer (bcc) is predicted with an accuracy of 95.61%. The deadliest skin cancer (mel) is predicted with an accuracy greater than 90%. Thus, the model performs well in diagnosing the most serious cases.

Table 2 Predictive accuracy by lesion classification
Full size table

5 Conclusions

A deep learning approach to diagnosing different types of skin lesions ranging from potentially deadly skin cancers to benign age spots was employed. Results indicate that an automated approach can be used to effectively diagnose the etiology of lesions, detecting skin cancers more accurately than human experts [9]. This is an important finding with high pragmatic value. An approach to skin cancer screening will greatly improve health outcomes for patients while reducing resource expenditures. For instance, patients will be able to obtain accurate preliminary diagnosis from their primary care physician without seeking out a specialist, rural patients would be able to obtain a diagnosis through telemedicine, and laboratories would likely see a reduction in the number of unnecessary biopsies needing to be processed. While further testing and refinement is required, we believe the current results can help healthcare providers to make more accurate decisions.