| 摘要: | 腎臟是人體重要的器官，然而當某些原因造成腎臟功能下降到一個程度時則需要洗腎，但是腎臟功能下降會形成血管鈣化，使血管失去彈性，最後提高死亡率。但是目前腹主動脈鈣化評分由專門的醫生來進行，所產生的問題為評分時醫生帶有主觀意識，導致可能對同一位患者的X光影像，但是不同位醫生評分或是同一位醫生在不同時間點評分，產生出分數不一致的情況，以及偶爾有為錯誤的腹主動脈節段評分的情況。本研究想要找到一個方法可以幫醫生更公正地去幫患者評估血管鈣化的程度。我們針對生理數據使用的模型為Logistic Regression、Ridge、SVM_Linear、K Nearest Neighbor、Random Forest和MLP；針對X光影像使用的模型有未pretrain的Resnet50、有pretrain的Resnet50和Vision Transformer。目前以生理數據訓練效果最好的是用RidgeCV挑選前三重要的因素訓練的Logistic Regression模型，準確率達67.22%；以影像訓練效果最好的是Vision Transformer，準確率達至54.17%，但是以原圖和做背景遮罩後的影像分別訓練的teachable machine模型準確率分別達75%和81.25%，表示將非觀察的區域當作背景遮住之後，可以有效的避免背景對於模型判斷的干擾。
;The kidneys are vital organs in the human body. When kidney function declines to a certain level, renal replacement therapy, such as dialysis, becomes necessary. Unfortunately, the decline of kidney function can lead to vascular calcification, causing blood vessels to lose their elasticity and ultimately increasing mortality risk. Currently, the assessment of abdominal aortic calcification is performed by specialized physicians. However, this approach has limitations as the scoring is subjective, resulting in inconsistent ratings between different physicians or even the same physician at different times. Occasionally, there may also be erroneous segment scoring of the abdominal aorta.
This study aims to find a method to help physicians assess the degree of vascular calcification in patients more objectively. We utilized models such as Logistic Regression, Ridge, SVM_Linear, K Nearest Neighbor, Random Forest, and MLP for physiological data. For X-ray images, we employed models including Resnet50 with and without pretraining and Vision Transformer. Currently, the best-performing model for physiological data is the Logistic Regression model trained using RidgeCV to select the top three important factors, achieving an accuracy of 67.22%. Regarding image analysis, the Vision Transformer model achieved the highest accuracy of 54.17%. However, when training the Teachable Machine model using the original images and images with background masking separately, the accuracies reached 75% and 81.25%, respectively. This indicates that masking the unobserved regions as background effectively avoids interference in the model’s judgment caused by the background. |
