Open-source images of melanoma and basal cell carcinoma (BCC) were acquired from the Hellenic Dermatological Atlas [4], the Dermatology Atlas [5], the Interactive Dermatology Atlas [6], and DermNet NZ [7]. Two image recognition models were trained, validated, and tested using methodology as described previously [8]. TensorFlow [9], an open-source software library by Google, was used as a deep-learning framework and was used to retrain Inception, version 3 (v3). Inception v3 is a deep convolutional neural network. This neural network consists of a hierarchy of multiple computational layers that each have an input and output. All layers except the final layer of this neural network are pretrained with more than 1.2 million images. The final layer of the neural network was retrained with the gathered dermatological images. During the retraining process, the neural network underwent both a training and validation step. In the training step, the inputted images were used to train the neural network. In the validation step, inputted naïve images were used to iteratively assess training accuracy [10].

After the model had been retrained (trained and validated), a user-inputted testing/assessment step was performed in which test images were inputted and the results were statistically analyzed. The program assessment output is expressed in terms of percentages of the probability of each of the dermatological manifestations for each testing image inputted. R software (R Foundation for Statistical Computing) [11] was used to perform the statistical analysis. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score were calculated for each dermatological manifestation. The F1 score is the harmonic average of the sensitivity and PPV (mean of the recall and precision).

The goal of each model was to differentiate between melanoma and BCC.

The first model was:

The second model was:

Area under the receiver operating characteristic (AUC) curves for melanoma and BCC were calculated to determine the performance of the two models.

When asked to differentiate between melanoma and BCC, the image recognition model trained and validated on images of light skin color had higher sensitivity, specificity, PPV, NPV, and F1 score than the image recognition model trained and validated on images of skin of color (Table 1).

In predicting melanoma, the image recognition model trained and validated on images of light skin color had a sensitivity of 0.60, specificity of 0.53, PPV of 0.56, NPV of 0.57, and F1 score of 0.58. On the other hand, in predicting melanoma, the same image recognition model trained and validated on images of skin of color had a sensitivity of 0.53, specificity of 0.47, PPV of 0.50, NPV of 0.50, and F1 score of 0.52.

In predicting BCC, the image recognition model trained and validated on images of light skin color had a sensitivity of 0.53, specificity of 0.60, PPV of 0.57, NPV of 0.56, and F1 score of 0.55. On the other hand, for prediction of BCC, the same image recognition model trained and validated on images of skin of color had a sensitivity of 0.47, specificity of 0.53, PPV of 0.50, NPV of 0.50, and F1 score of 0.48.

The average AUC for the two light skin color image recognition models was 0.598, compared to 0.500 (values point out the difference) for the skin of color image recognition models (Table 1 and Figure 1).