Design and validation of a new machine-learning-based diagnostic tool for the differentiation of dermatoscopic skin cancer images

Abstract

Background: Skin cancer is the most common cancer in the United States. Current estimates are that one in five Americans will develop skin cancer in their lifetime. A skin cancer diagnosis is challenging for dermatologists requiring a biopsy from the lesion and histopathological examinations. In this article, we used the HAM10000 dataset to develop a web application that classifies skin cancer lesions.

Method: This article presents a methodological approach that utilizes dermoscopy images from the HAM10000 dataset, a collection of 10015 dermatoscopic images collected over 20 years from two different sites, to improve the diagnosis of pigmented skin lesions. The study design involves image pre-processing, which includes labelling, resizing, and data augmentation techniques to increase the instances of the dataset. Transfer learning, a machine learning technique, was used to create a model architecture that includes EfficientNET-B1, a variant of the baseline model EfficientNET-B0, with a global average pooling 2D layer and a softmax layer with 7 nodes added on top. The results of the study offer a promising method for dermatologists to improve their diagnosis of pigmented skin lesions.

Results: The model performs best in detecting melanocytic nevi lesions with an F1 score of 0.93. The F1 score for Actinic Keratosis, Basal Cell Carcinoma, Benign Keratosis, Dermatofibroma, Melanoma, and Vascular lesions was consecutively 0.63, 0.72, 0.70, 0.54, 0.58, and 0.80.

Conclusions: We classified seven distinct skin lesions in the HAM10000 dataset with an EfficientNet model reaching an accuracy of 84.3%, which provides a promising outlook for further development of more accurate models.

Fig 1. Data augmentation.

Illustration of how data augmentation algorithm works. A, 25 intact random images from the dataset; B, C, and D, random data augmentations applied to that same 25 images, which would then feed to the model.

Fig 2. The model’s architecture.

The 340 layers of the EfficientNET-B1 is collapsed.

Fig 3. Learning curves.

The alterations in loss and accuracy of the model’s predictions per epoch is illustrated.

Fig 4. Dataset description.

Fig 5. Frequency of each type of skin cancer lesions in the dataset.

Fig 6. ANOVA test to compare mean age between different classes of skin cancer and gender.

Dashed lines indicate Mean and ± one standard deviation interval from Mean. Dots indicate outliners.

Fig 7. Model metrics for each class.

AKIEC: Actinic Keratosis; BCC: Basal Cell Carcinoma; BKL: Benign Keratosis; DF: Dermatofibroma; MEL: Melanoma; NV: Melanocytic Nevi; VASC: Vascular Lesions.

Fig 8. The top 20 most wrong predictions.

Actual: the true label of each image; pred: the model’s predicted class; prob: the probability inferred to the predicted class by the model.