基於三切面ResUNet融合架構與生成式人工智慧的肝臟腫瘤自動分割

Abstract

肝癌是全球第二致命的癌症，治療方式包括手術切除和血管灌流治療等。肝臟檢測通常依賴電腦斷層掃描（CT）影像，但手動標記影像耗時且受臨床醫師經驗影響。另外，臨床上許多重要的治療方法都需要精準的計算肝葉體積。因此，本研究首先使用兩個二維模型：ResUNet 以及ResNet50 + FCN16 進行自動化左右肝分割以及體積計算，以提高醫療影像分割效率、減輕臨床醫師負擔。 此外，三維分割模型相較二維模型表現較佳，但耗費大量運算資源。本研究提出三切面 ResUNet 融合架構構來達到自動化肝臟腫瘤分割。該架構藉由融合三切面的二維模型預測結果，來克服過去二維模型在單一切面分割病灶的劣勢。但因為醫療影像切片之間的距離通常比切片內的解析度來得高，因此冠狀面以及矢狀面的影像解析度會因此較低。為了解決三維影像資料在 z 軸的解析度不足的問題，提出一資料擴增做法，採用生成式人工智慧（GAI）策略，利用神經網模型從相鄰切片生成合成切片，從而提高矢狀面和冠狀面的影像解析度。與其他基於幾何轉換的擴增做法不同，該擴增做法除了有效增加訓練樣本切片，也提高單一醫療立體影像內的資訊。 最後在全肝體積估算上，使用 ResUNet 達到了 0.6% 的差異比；使用 ResNet50 + FCN16 達到了 2.7% 的差異比。另外，透過所提出的資料擴增做法，顯著地提高冠狀面以及矢狀面的分割模型成效，也有助於提高三切面ResUNet 融合架構的分割表現。在肝腫瘤分割任務上取得了 74.4% 的 Dice 分數。除此之外，我們進一步的展示了提出的融合架構在判斷邊緣腫瘤像素的精確性上有所提升。

Liver cancer is the second leading cause of cancer-related mortality worldwide. Standard treatments, including surgical resection and intra-arterial infusion, heavily depend on diagnostic imaging, particularly computed tomography (CT) scans. However, manual annotation of CT images is both time-consuming and highly dependent on the clinician’s experience. Furthermore, many critical clinical treatments require precise volumetric calculations of liver lobes. Therefore, this study first utilizes two-dimensional models: ResUNet and ResNet50 combined with FCN16, for the automated segmentation and volumetric calculation of the left and right liver lobes. This approach can enhance labeling efficiency and alleviate the burden on clinicians. While three-dimensional segmentation models generally perform better than their two-dimensional counterparts, they demand significant computational resources. To address this, we propose a Tri-plane ResUNet Model Fusion architecture for the automated segmentation of liver tumors. This architecture leverages the predictive capabilities of fused three-plane models, effectively addressing the limitations associated with previous two-dimensional models that segmented lesions on a single plane. By integrating the segmentation results from axial, coronal, and sagittal planes, this fusion approach overcomes the shortcomings of single-plane segmentation. A significant challenge in medical imaging is the lower resolution in the z-axis compared to the in-plane resolution, which affects the clarity of sagittal and coronal images. To address this issue, we introduce a generative artificial intelligence (GAI) strategy that employs neural network models to generate synthetic slices from adjacent ones, thereby enhancing the resolution and quality of images in the sagittal and coronal planes. This approach not only augments the training dataset with realistic synthetic slices but also enriches the information content within a single volumetric medical image, surpassing traditional geometric transformation-based augmentation methods. For the overall liver volume estimation, ResUNet achieved a difference ratio of 0.6%, and ResNet50 combined with FCN16 achieved a difference ratio of 2.7%. The proposed data augmentation technique significantly improves the performance of segmentation models for the coronal and sagittal planes, contributing to the overall effectiveness of the Tri-plane ResUNet Model Fusion architecture. In liver tumor segmentation tasks, we achieved a Dice score of 74.4%. Additionally, we demonstrate that the fusion architecture notably improves the accuracy of detecting tumor boundary pixels, highlighting its potential for clinical applications.