PDF(Pubmed)
Abstract：
This study has used machine learning algorithms to develop a predictive model for differentiating between dermoscopic images of basal cell carcinoma (BCC) and actinic keratosis (AK). We compiled a total of 904 dermoscopic images from two sources - the public dataset (HAM10000) and our proprietary dataset from the First Affiliated Hospital of Dalian Medical University (DAYISET 1) - and subsequently categorised these images into four distinct cohorts. The study developed a deep learning model for quantitative analysis of image features and integrated 15 machine learning algorithms, generating 207 algorithmic combinations through random combinations and cross-validation. The final predictive model, formed by integrating XGBoost with Lasso regression, exhibited effective performance in the differential diagnosis of BCC and AK. The model demonstrated high sensitivity in the training set and maintained stable performance in three validation sets. The area under the curve (AUC) value reached 1.000 in the training set and an average of 0.695 in the validation sets. The study concludes that the constructed discriminative diagnostic model based on machine learning algorithms has excellent predictive capabilities that could enhance clinical decision-making efficiency, reduce unnecessary biopsies, and provide valuable guidance for further treatment.
摘要：
这项研究使用机器学习算法来开发预测模型，以区分基底细胞癌（BCC）和光化性角化病（AK）的皮肤镜图像。我们从两个来源-公共数据集（HAM10000）和大连医科大学第一附属医院的专有数据集（DAYISET1）-共编辑了904张皮肤镜图像，随后将这些图像分为四个不同的队列。该研究开发了用于定量分析图像特征的深度学习模型，并集成了15种机器学习算法，通过随机组合和交叉验证生成207个算法组合。最终的预测模型，通过将XGBoost与Lasso回归集成而形成，在BCC和AK的鉴别诊断中表现出有效的表现。该模型在训练集中表现出高灵敏度，并在三个验证集中保持稳定的性能。训练集中的曲线下面积(AUC)值达到1.000,验证集中的平均值为0.695。研究结论:构建的基于机器学习算法的判别诊断模型具有良好的预测能力,可以提高临床决策效率,减少不必要的活检，为进一步治疗提供有价值的指导。
