Computer Aided Diagnosis of Melanoma
Computer Aided Diagnosis of Melanoma
David Xiang Thomas Jefferson High School for Science and Technology
This article was originally published in the 2021 print edition of Teknos Science Journal.
I counted twenty evenly spread across my arms, chest, and legs. The small brown moles had never bothered me before, but after my health and physical education teacher showed me a photograph of melanoma, I couldn’t help but worry. What if the innocent moles I used to call freckles are actually cancerous melanoma cells? That night, I spent an hour examining each of my moles in detail, but this proved harder than expected. Some moles were barely darker than my skin tone, others were hidden on my back and out of sight. I eventually gave up on my self-examination. Since none of my moles have grown, I think I am lucky enough to not have melanoma. Regardless, melanoma has become more widespread throughout the last few decades [5]. Therefore, it has become necessary to find a way to self-examine possible melanoma moles from home. Through the Computer Systems lab at Thomas Jefferson High School for Science and Technology, I aim to create a mobile application capable of determining if a mole is cancerous using image recognition and machine learning.
Melanoma, a type of skin cancer in which pigment-producing melanocyte skin cells turn cancerous, is traditionally diagnosed by experienced dermatologists and oncologists. Unfortunately, this is often inconvenient. The average person can have anywhere between ten to forty moles, and it’s both expensive and troublesome to visit a doctor every time a new mole shows up. In addition, the difference between a benign mole and melanoma in its early stages is almost nonexistent. In the past two decades, an alternative to human diagnosis of melanoma arose. New image recognition and machine learning techniques have allowed computers to accurately diagnose melanoma. Although less accurate than an experienced doctor, mobile apps trained to discern benign moles from malignant melanoma are easily accessible online. Recently, the development and popularity of deep learning further increases the accuracy of computer aided melanoma diagnosis [4].
To train a machine learning model to be able to tell the difference between benign moles and melanoma, researchers need a large image database of moles. Phillips et al. (2020) used a subset of the ISIC archive, which contains more than 7000 images of moles, to train an artificial intelligence neural network named Deep Ensemble for Recognition of Melanoma (DERM). They filtered the ISIC archive, excluding images of children’s moles because skin cancer is rare in children, meaning their inclusion would harm diagnostic accuracy. In addition to the ISIC database, melanoma researcher Michael Phillips (personal communication, Feb. 14, 2021) was able to collect dermoscopic images from local clinics. This may be hard to replicate for my project however due to patient privacy rights. DERM performed better than clinicians that examined the same test images. DERM’s diagnostic accuracy was calculated using the receiver operating characteristic curve. With an area under the curve of 0.93, DERM had a better diagnostic accuracy than physicians who had an area under the curve of 0.83.
Convolutional neural networks have become popular in the field of computer-aided diagnosis because of their ability to classify images. Haenssle et al. (2018) modified and trained a version of Google’s Inception v4 convolutional neural network to differentiate between images of melanoma and harmless moles. In order to improve the accuracy of their neural network, the researchers inputted patient information into the neural network. Haenssle et al. (2018) found that including patient information, such as sex, age, and body site of mole, significantly increased the accuracy of the convolutional neural network. In my research, I am looking to use my own convolutional neural network to perform a similar function. Haenssle et al. (2018) then gave 100 test images of melanoma and cancerless moles to 58 dermatologists to compare the neural network’s diagnostic accuracy to the doctors’s diagnostic accuracy. With a diagnostic accuracy of 82.5%, the convolutional neural network outperformed the dermatologists, who achieved a diagnostic accuracy of 75.7%.
Combining both human and artificial intelligence can improve diagnostic accuracy of melanoma. Heckler, Utikal, Enk, and Hauschild (2019) trained a convolutional neural network using 11,000 images of moles derived from the same ISIC database that Phillips et al. used. The convolutional neural network used binary classification to classify the image as either benign or malignant. Then, Heckler et al. (2019) sent an electronic survey to 13 different dermatologists and had them classify a sample of moles as melanoma, nevus meaning a birthmark, or a different form of skin cancer. The dermatologists also recorded their confidence for each diagnosis. Then, the researchers fused both classification results to create a more accurate classifier. The accuracy of the convolutional neural network was 81.59%.This was improved by 1.36% after fusing with the human classification. In my project, I am attempting to replicate the human classification process by reproducing the steps a dermatologist would go through when presenting a possible melanoma mole. This classification will then be fused with my project’s convolutional neural network much like how Heckler et al. (2019) fused their classifications.
While the accuracy of computer aided diagnosis will continue to improve throughout the next few years, there are many obstacles to overcome before computers can diagnose melanoma with near perfect accuracy. Future research in this field will involve increasing the database of melanoma images, so machine learning models can train from a more robust set of images. Furthermore, new image recognition techniques will improve the accuracy of the current machine learning models. For now, I hope my project will be able to build upon the successes of other melanoma identification applications, and eventually save lives by accurately diagnosing melanoma in its early stages. My project will look to incorporate some of the current image recognition and computer diagnostic techniques to classify a picture of a mole as either melanoma or nevus. If my project achieves high diagnostic accuracy, it could find a place within doctor’s offices and assist oncologists and dermatologists in their diagnoses.
References
[1] Grady, D. (2020, January 6). A.I. comes to the operating room. The New York Times. https://www.nytimes.com/2020/01/06/health/artificial-intelligence-brain-cancer.html
[2] Haenssle, H.A., Fink, C., Schneiderbauer, R., Toberer, F., Buhl, T., Blum, A., Kalloo, A., Hassen, A. B. H., Thomas, L., Enk, A., Uhlmann, L., Alt, C., Arenbergerova, M., Bakos, R., Baltzer, A., Bertlich, I., Blum, A., Bokor-Billmann, T., Bowling, J., . . . Zalaudek, I. (2018). Man against machine: Diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists. Annals of Oncology, 29(8), 1836-1842. https://doi.org/10.1093/annonc/mdy166
[3] Heckler, A., Utikal, J., Enk, A., & Hauschild, A. (2019). Superior skin cancer classification by the combination of human and artificial intelligence. European Journal of Cancer, 120, 114-121. https://doi.org/10.1016/j.ejca.2019.07.019
[4] Hosny, A., Parmar, C., & Quackenbush, J. (2018). Artificial intelligence in radiology. Nature Reviews Cancer, 500-510. https://doi.org/10.1038/s41568-018-0016-5
[5] Nikolaou, V., & Stratigos, A.J. (2014). Emerging trends in the epidemiology of melanoma. British Journal of Dermatology, 170(1), 11-19. https://doi.org/10.1111/bjd.12492
[6] Phillips, M., Greenhalgh, J., Marsden, H., & Palamaras, I. (2020). Detection of malignant melanoma using artificial intelligence: An observational study of diagnostic accuracy. Dermartol Pract Concept. https://doi.org/10.5826/dpc.1001a11
