Related Experiment Video
Updated: Jun 12, 2025
Combining Reflectance Confocal Microscopy with Optical Coherence Tomography for Noninvasive Diagnosis of Skin Cancers via Image Acquisition
Published on: August 18, 2022
Skin Cancer Detection in Diverse Skin Tones by Machine Learning Combining Audio and Visual Convolutional Neural
Bruce N Walker1, Travis Wayne Blalock2, Rebecca Leibowitz2
1School of Psychology and School of Interactive Computing, Georgia Institute of Technology, Atlanta, Georgia, USA.
Introduction:
Skin cancer (SC) is common in fair skin (FS) at a 1:5 lifetime incidence for nonmelanoma skin cancer. In order to assist clinicians' decisions, a risk intervention technology was developed, which combines a dual-mode machine learning of visual and sonified (pixel to sound) data. The addition of an audio technology enhances malignant features of lesions, increases sensitivity and was previously validated under a prospective clinical setting in FS. In dark skin (DS), although rare by a 10-30 factor, skin cancer is diagnosed at more advanced stages resulting in a delayed diagnosis and affecting life quality and expectancy. It is known as well that SC diagnostic accuracy by machine learning in DS is decreased as compared to FS. The present study tests the use of sonification aided by artificial intelligence algorithms to compare diagnostics of different skin tones.
Methodology:
Biopsy-validated smartphone images were diagnosed in a retrospective study by a dual audio-visual convoluted neural network. A total of 60 Fitzpatrick I-III were compared to 72 Fitzpatrick IV-VI. A dichotomous diagnostic output, either malignant or benign, was assessed for sensitivity, specificity and area under the curves (AUCs) for the receiver operating characteristic (ROC) curve.
Results:
ROC curve analytics indicated an AUC of 0.858 (95% CI: 0.795-0.921) and 0.856 (95% CI: 0.759-0.953) for fair and DS (p = NS). Sensitivity of Fitzpatrick I-III skin and Fitzpatrick IV-VI were 84.4% (71.8-96.9) and 79.6% (63.4-93.8), respectively (p = NS). Specificity of Fitzpatrick I-III skin and Fitzpatrick IV-VI were 84.2% (72.6-95.8) and 85.3% (73.4-97.2), respectively (p = NS). The positive predictive and negative predictive values as well as accuracy (0.817 vs. 0.847) were all within the same range (p = NS).
Conclusions:
The results demonstrate that the dual-modality classifier identifies skin cancer of FS and DS similarly well. Sonification of malignant signs of a skin lesion demonstrates promising results, even with smartphone images, which should be considered as a tool to achieve more effective and accessible healthcare.
