人工智慧在醫學影像上之應用與模型參數分析

Abstract

本文中我們使用全域式光學同調斷層掃描得到的三圍老鼠皮膚病變影像作為資料集，並利用機器學習的原理，建構出能用於皮膚病變診斷的分類模型，評估各個模型後對模型架構進行修改，選擇出診斷準確率最好的CNN模型，並且對模型的感受野參數進行更深入的研究，探討感受野的變化對模型準確率的影響，挑選出最適合的參數數值與模型架構，最終得到81.43%的準確率，可知本模型在皮膚癌的臨床診斷上具備一定的可行性。

In this study, we utilized three-dimensional skin lesion images obtained through full-field optical coherence tomography (FF-OCT) as our dataset. Leveraging the principles of machine learning, we constructed a classification model for diagnosing skin lesions. After evaluating various models, we made modifications to the model structure. The CNN model with the highest accuracy was selected. Furthermore, we conducted an in-depth investigation into the parameters of the model. We explored the impact of changes in receptive field on model accuracy, identifying the most suitable parameter values and model structure. As a result, we achieved an accuracy of 81.43%. This suggests that our model demonstrates a certain level of feasibility for clinical diagnosis of skin lesions and skin cancer.