基於深度學習的醫學影像語義分割案例研究

Case Studies of Medical Image Semantic Segmentation Based on Deep Learning

張逸宗(Yi-Zong Zhang)

指導教授 ： 周信宏

國立暨南國際大學/科技學院/資訊工程學系/碩士(2023年)

https://doi.org/10.6837/ncnu202300008

摘要

近年來深度學習技術快速發展，各種先進的深度學習模型接連被提出，並應用於不同專業領域。醫療影像辨識是其中一個最常見的應用。早期的醫療影像辨識通常只是分類(classification)，例如判斷有病或沒病。然而為什麼做出這樣的判斷結果並沒有說明，使得這樣的模型無法實際在臨床上使用。近年醫療影像辨識的發展方向是朝向可解釋性發展，透過語義分割(semantic segmentation)將影像中的病徵區域標記出來，讓醫生能夠透過肉眼驗證答案。常見的醫療影像包括攝影機照片、X光(X-ray)、斷層掃描(CT)、磁振造影(MRI)和超音波(Ultrasound)等。本研究針對灰階的超音波影像和彩色的眼底鏡影像進行病徵的語義分割研究。
　　我們在超音波影像中使用了U-Net、Res-UNet及UNet++進行心包膜積液和水腎症積水分割，並比較了三者的效能。此外我們還設計了AI協同標記方法，使病症標記能夠排除人眼疏漏，減少模型學習矛盾，以提升準確率。我們從101名有中/重度心包膜積液的病人中提取了2,387張影像；從97名水腎病人中提取了3,462張影像，並分別提取了2,753張和1,628張正常影像進行兩項任務的訓練和驗證。最後利用後處理方法使辨識結果能更貼近臨床實務。實驗結果顯示：心包膜積液的準確率達到96.2%；水腎症的準確率達到94.6%。引入臨床應用將能有效提升急重症處置品質。
　　U-Net系列模型透過融合不同深度的特徵來推斷最終分割結果。該模型在處理只有單一目標的灰階超音波影像時表現良好，但是當任務變為彩色影像、背景複雜、多目標且尺寸變化極大的眼底鏡糖尿病視網膜病變時，則無法達到同等效果。因此本研究使用Transformer設計了一種具有特徵深度和空間之雙注意力機制模型，並在眼底鏡公開資料集IDRiD上進行評估。其中微血管瘤的Precision-Recall曲線下面積(Area under Curve of Precision-Recall, AUCPR)為51.07%，高於目前最佳方法的50.99%。在臨床實務中，將能協助醫師找出微小的早期病徵，早期治療以降低失明的風險。

關鍵字

超音波 ； 心包膜積液 ； 水腎症 ； 眼底鏡 ； 糖尿病視網膜病變 ； 語意分割 ； 深度學習 ； Transformer

並列摘要

In recent years, deep learning technology has developed rapidly, and various advanced models have been proposed and applied to different professional fields. Medical image recognition is one of the most common applications. Early medical image recognition is usually just classification, such as diagnosing disease or absence. However, why such a result was made is not explained, making such a model impractical for clinical use. In recent years, the development direction of medical image recognition has been towards interpretability, marking the disease regions in images through semantic segmentation, and allowing doctors to verify the results through visual inspection. Common medical images include camera photos, X-rays, CT scans, MRIs, and ultrasounds. This study focuses on the semantic segmentation of grayscale ultrasound images and color fundus images.
In this study, we utilized U-Net, Res-UNet, and UNet++ for the segmentation of pericardial effusion and hydronephrosis in ultrasonic images and compared the performance of the three models. Additionally, we designed an AI-assisted annotation method to eliminate human annotation errors and reduce model learning conflicts, thereby improving accuracy. We extracted 2,387 images from 101 patients with moderate to severe pericardial effusion and 3,462 images from 97 hydronephrosis patients, and 2,753 and 1,628 normal images respectively for training and validation of the two tasks. Finally, post-processing methods were used to better align the recognition results with clinical practice. The experimental results showed that the accuracy for pericardial effusion reaches 96.2% and for hydronephrosis reaches 94.6%. Introducing our models to clinical applications will effectively improve the quality of emergency treatment.
The U-Net series models predict the final segmentation result by fusion features of different depths. The model performs well in processing grayscale ultrasound images with a single target but fails to deliver equivalent results for more complex tasks such as color images, complicated backgrounds, multi-target, and large-size variations in diabetic retinopathy fundus images. To address this challenge, we propose a Transformer-based model with a dual attention mechanism that considers both feature depth and spatial information. The proposed model was evaluated on the IDRiD public dataset for retinal images. The model achieved an Area under the Curve of Precision-Recall (AUCPR) of 51.07% for microaneurysms, surpassing the current best method of 50.99%. This can assist physicians in identifying tiny early symptoms, facilitating early treatment, and reducing the risk of vision loss in clinical practices.

並列關鍵字

Ultrasound ； Pericardial Effusion ； Hydronephrosis ； Ophthalmoscope ； Diabetic Retinopathy ； Semantic Segmentation ； Deep Learning ； Transformer

參考文獻

參考文獻

[1] F. E.-Z. A. El-Gamal, M. Elmogy and A. Atwan, "Current trends in medical image registration and fusion," in Egyptian Informatics Journal, 17(1), 99-124, 2016.

[2] C. Chen, C. Qin, H. Qiu, G. Tarroni, J. Duan, W. Bai and D. Rueckert, "Deep learning for cardiac image segmentation: a review," in Frontiers in Cardiovascular Medicine, 7, 25, 2020.

[3] A. Pérez-Casares, S. Cesar, L. Brunet-Garcia and J. Sanchez-de-Toledo, "Echocardiographic evaluation of pericardial effusion and cardiac tamponade," in Frontiers in pediatrics, 5, 79, 2017.

[4] S. A. Alshoabi, D. S. Alhamodi, M. A. Alhammadi and A. F. Alshamrani, "Etiology of Hydronephrosis in adults and children: Ultrasonographic Assessment in 233 patients," in Pakistan Journal of Medical Sciences, 37(5), 1326, 2021.