- 學位論文
深度學習於自動乳房超音波三維腫瘤偵測
3D Tumor Detection for Automated Breast Ultrasound using Deep Learning Approaches
若您是本文的作者，可授權文章由華藝線上圖書館中協助推廣。
摘要
乳癌是全球女性癌症死亡的主要原因之一。自動乳房超音波(Automated Breast Ultrasound, ABUS)已成為一種重要的篩檢方式，尤其特別適合乳房組織緻密的女性。然而，ABUS影像解讀需要放射科醫師檢視數百張影像切片，導致診斷過程耗時且容易受到解讀者的疲勞影響。因此，本論文提出一種創新的電腦輔助偵測(Computer-aided Detection, CADe)系統，該系統結合了YOLO3D-ABUS，是一種專門設計用於解決ABUS影像解讀的單階段三維卷積神經網路。 本研究提出的CADe系統包含三個主要部分：標準化的體積資料預處理、使用YOLO3D-ABUS進行體積腫瘤偵測，以及結果優化的策略性後處理。我們的腫瘤偵測方法將YOLOv8的偵測框架從原有的二維實作擴展為完全三維架構，直接處理ABUS的三維影像資料，從而消除了二維切片式偵測方法所伴隨的空間連續性問題。此架構整合了MedNeXt的元素，以加強醫學影像環境中的特徵提取能力。為進一步優化模型在乳房腫瘤定位方面的表現，我們的一項關鍵創新是提出scaled 3-D DIOU，該函數將原始DIOU公式擴展為適合考量乳腺腫瘤在三維空間中的體積特性。 我們的系統在使用包含523個確診腫瘤的258個ABUS體積資料進行的全面測試中展現出卓越性能，結果顯示在90%、95%、98%和99%的敏感度下，對應的平均每次掃描假陽性數量分別達到0.65、1.44、5.01和7.54，並且自由反應接收者操作特徵（Free-response Receiver Operating Characteristic , FROC）曲線下標準化部分面積達到0.956。此系統相較於現有方法更能有效處理每個ABUS體積資料，特別是在偵測傳統方法中常常忽略的較小型腫瘤。 本研究對自動化乳癌偵測領域做出貢獻，透過更精確且高效的乳癌篩檢，有望改善臨床工作流程和病患預後。
並列摘要
Breast cancer represents one of the leading causes of cancer-related mortality among women worldwide. Automated breast ultrasound (ABUS) has emerged as a valuable screening modality, particularly beneficial for women with dense breast tissue. However, ABUS interpretation requires radiologists to review hundreds of image slices, creating a time-consuming process prone to interpreter fatigue. Therefore, this thesis presents a novel computer-aided detection (CADe) system incorporating YOLO3D-ABUS, a specialized one-stage 3-D convolutional neural network designed to address the challenge of interpreting ABUS volumes. The proposed CADe system comprises three main components: volume preprocessing for standardization, volumetric tumor detection using YOLO3D-ABUS, and strategic postprocessing for result refinement. Our tumor detection approach extends the detection framework of YOLOv8 from its original 2-D implementation to a fully 3-D architecture that processes ABUS volumes directly in their three-dimensional form, eliminating the spatial continuity problems associated with 2-D slice-by-slice detection methods. This architecture integrates specialized elements from MedNeXt for enhanced feature extraction in medical imaging contexts. To further optimize the model's performance in localizing breast tumors, a key innovation is our scaled 3-D distance-intersection over union (scaled 3-D DIOU), which extends the original DIOU formulation to properly account for the volumetric nature of breast tumors in three-dimensional space. Our system demonstrated exceptional performance in comprehensive testing using a dataset of 258 ABUS volumes containing 523 confirmed tumors, achieving sensitivities of 90%, 95%, 98%, and 99% with corresponding false positives per pass of 0.65, 1.44, 5.01, and 7.54, respectively, and a normalized partial area under free-response receiver operating characteristic (FROC) curve of 0.956. The system processes each ABUS volume efficiently, representing a significant improvement over existing methods, particularly for detecting smaller tumors that are often missed by conventional approaches. This work contributes to the field of automated breast cancer detection with potential implications for improving clinical workflows and patient outcomes through more accurate and efficient breast cancer screening.
參考文獻
[1] R. L. Siegel, A. N. Giaquinto, and A. Jemal, "Cancer statistics, 2024," CA: a cancer journal for clinicians, vol. 74, no. 1, pp. 12-49, 2024.
[2] A. Vourtsis and W. A. Berg, "Breast density implications and supplemental screening," European radiology, vol. 29, pp. 1762-1777, 2019.
[3] B. Wilczek, H. E. Wilczek, L. Rasouliyan, and K. Leifland, "Adding 3D automated breast ultrasound to mammography screening in women with heterogeneously and extremely dense breasts: Report from a hospital-based, high-volume, single-center breast cancer screening program," European journal of radiology, vol. 85, no. 9, pp. 1554-1563, 2016.
[4] P. Skaane, R. Gullien, E. B. Eben, M. Sandhaug, R. Schulz-Wendtland, and F. Stoeblen, "Interpretation of automated breast ultrasound (ABUS) with and without knowledge of mammography: a reader performance study," Acta Radiologica, vol. 56, no. 4, pp. 404-412, 2015.
[5] D. G. Lowe, "Distinctive image features from scale-invariant keypoints," International journal of computer vision, vol. 60, pp. 91-110, 2004.