Ofira Zloto, Ofir Fogel, Guy Ben Simon, Mordechai Rosner, Vicktoria Vishnevskia-Dai, Avner Hostovsky, Eyal Klang
{"title":"Computer-aided diagnosis of eyelid skin tumors using machine learning.","authors":"Ofira Zloto, Ofir Fogel, Guy Ben Simon, Mordechai Rosner, Vicktoria Vishnevskia-Dai, Avner Hostovsky, Eyal Klang","doi":"10.1016/j.jcjo.2024.07.015","DOIUrl":null,"url":null,"abstract":"<p><strong>Objective: </strong>To develop an automated, new framework based on machine learning to diagnose malignant eyelid skin tumors.</p><p><strong>Methods: </strong>This study used eyelid lesion images from Sheba Medical Center, a large tertiary center in Israel. Before model training, we pretrained our models on the International Skin Imaging Collaboration (ISIC) 2019 dataset consisting of 25,332 images. The proprietary eyelid data set was then used for fine-tuning. The data set contained multiple images per patient, aiming to classify malignant lesions in comparison to benign counterparts.</p><p><strong>Results: </strong>The analyzed data set consisted of images representing both benign and malignant eyelid lesions. For the benign category, a total of 373 images were sourced. By comparison, for the malignant category, 186 images were sourced. For the final model, at sensitivity of 93.8% (95% CI 80.0-100.0%), the model has a corresponding specificity of 73.7% (95% CI 60.0-87.1%). To further understand the decision-making process of our model, we employed heatmap visualization techniques, specifically gradient-weighted Class Activation Mapping.</p><p><strong>Discussion: </strong>This study introduces a dependable model-aided diagnostic technology for assessing eyelid skin lesions. The model demonstrated accuracy comparable to human evaluation, effectively determining whether a lesion raises a high suspicion of malignancy or is benign. Such a model has the potential to alleviate the burden on the health care system, particularly benefiting rural areas, and enhancing the efficiency of clinicians and overall health care.</p>","PeriodicalId":9606,"journal":{"name":"Canadian journal of ophthalmology. Journal canadien d'ophtalmologie","volume":" ","pages":""},"PeriodicalIF":3.3000,"publicationDate":"2024-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Canadian journal of ophthalmology. Journal canadien d'ophtalmologie","FirstCategoryId":"3","ListUrlMain":"https://doi.org/10.1016/j.jcjo.2024.07.015","RegionNum":4,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"OPHTHALMOLOGY","Score":null,"Total":0}
引用次数: 0
Abstract
Objective: To develop an automated, new framework based on machine learning to diagnose malignant eyelid skin tumors.
Methods: This study used eyelid lesion images from Sheba Medical Center, a large tertiary center in Israel. Before model training, we pretrained our models on the International Skin Imaging Collaboration (ISIC) 2019 dataset consisting of 25,332 images. The proprietary eyelid data set was then used for fine-tuning. The data set contained multiple images per patient, aiming to classify malignant lesions in comparison to benign counterparts.
Results: The analyzed data set consisted of images representing both benign and malignant eyelid lesions. For the benign category, a total of 373 images were sourced. By comparison, for the malignant category, 186 images were sourced. For the final model, at sensitivity of 93.8% (95% CI 80.0-100.0%), the model has a corresponding specificity of 73.7% (95% CI 60.0-87.1%). To further understand the decision-making process of our model, we employed heatmap visualization techniques, specifically gradient-weighted Class Activation Mapping.
Discussion: This study introduces a dependable model-aided diagnostic technology for assessing eyelid skin lesions. The model demonstrated accuracy comparable to human evaluation, effectively determining whether a lesion raises a high suspicion of malignancy or is benign. Such a model has the potential to alleviate the burden on the health care system, particularly benefiting rural areas, and enhancing the efficiency of clinicians and overall health care.
期刊介绍:
Official journal of the Canadian Ophthalmological Society.
The Canadian Journal of Ophthalmology (CJO) is the official journal of the Canadian Ophthalmological Society and is committed to timely publication of original, peer-reviewed ophthalmology and vision science articles.