https://www.ncbi.nlm.nih.gov/pubmed/30446202
How I learned to stop worrying and love machine learning.
Author information
- 1
- University of Connecticut School of Medicine, Farmington, Connecticut, USA.
- 2
- Department of Dermatology, University of Connecticut School of Medicine, Farmington, Connecticut, USA.
- 3
- Department of Dermatology/Cutaneous Oncology and VA Palo Alto Health Care System, Stanford University Medical Center and Cancer Institute, Farmington, Connecticut, USA.
- 4
- Department of Dermatology, University of Connecticut School of Medicine, Farmington, Connecticut, USA. Electronic address: grant@uchc.edu.
Abstract
Artificial intelligence and its machine learning (ML) capabilities are very promising technologies for dermatology and other visually oriented fields due to their power in pattern recognition. Understandably, many physicians distrust replacing clinical finesse with unsupervised computer programs. We describe convolutional neural networks and discuss how this method of ML will impact the field of dermatology. ML is a form of artificial intelligence well suited for pattern recognition in visual applications. Many dermatologists are wary of such unsupervised algorithms and their future implications.
Copyright © 2018 Elsevier Inc. All rights reserved.
- PMID:
- 30446202
- DOI:
- 10.1016/j.clindermatol.2018.06.003
겁먹은 피부과 의사가 ML 을 대처하는 마음의 자세....
이런 이야기도 article 에 올려주는 친절..... ..
심약자들 을 위한, reassurance 처방 같은 (도움되는?) 어드바이스들 ....
Fear #1: I don’t understand ML.
A convolutional neural network (CNN) (Figure 1) is a type of ML that models the connections of biological neurons. The machine must be trained using images (input) labeled with the correct diagnosis (output). Images are analyzed pixel by pixel, serving as “neurons” that fire only if surpassing a certain threshold value for a specific image feature (a layer). Much like how real neuronal connections are constantly pruned based on experiential outcomes, CNNs select for connections between layers based on diagnostic performance – creating an association between features that best correlate to a diagnosis. Unfortunately, there is no way to know what features or threshold levels are being selected for (i.e. the “black box” of ML). 1
Fear #2: Machines will replace me.
Esteva et al. trained a Google-based CNN to diagnose melanomas and keratinocytic carcinomas with equal or superior success compared to 21 dermatologists, though not in a real-world clinical setting. 2 Another study comparing ML algorithms for characterizing pigmented lesions demonstrated superiority over 8 dermatologists in diagnosing melanoma. 3 While this seems threatening to the dermatologist’s diagnostic skillset, it is important to remember than ML provides only a probability score of various diagnoses – not a diagnosis, and certainly not management of the patient. We contend that ML will supplement and improve, not replace, a dermatologist’s capabilities. A recent metaanalysis in melanoma screening showed that current ML diagnostic success parallels that of dermoscopy, 4 and therefore may have more utility for non-dermatologists, particularly those attempting to triage skin lesions. This technology could potentially shift the full body screening to the primary care level, improving earlier detection of skin cancers and increasing appropriate referrals.
Fear #3: I am liable for machine errors.
Medical malpractice in the United States holds physicians liable when their care deviates from accepted standards to the point of negligence. An argument could be made that when an AI-provided diagnosis is superior to that of a dermatologist, deferring to AI would become best practice. However, this would involve rigorous, prospective clinical validation of AI algorithms to ensure diagnostic accuracy and superiority over specialist interpretation. Presently, deferring to AI is not an accepted standard and there is no legal precedent involving this technology. The Food and Drug Administration has provided vague guidelines that suggest clinical decision-making software with hidden decision processes will be classified and regulated as medical devices.5 However, no definitive regulation exists. We suggest physicians exercise judgement and treat AI-provided diagnoses as AI-guided diagnoses until clinical trial data are available.
