本研究探討了基於深度學習的胰臟分割方法。隨著醫學影像技術的不斷進步，胰臟分割在醫學診斷中的應用變得越來越重要。由於胰臟位於腹腔深處，其紋理特徵複雜且變異性大，使得自動分割面臨巨大挑戰。 本研究提出了一種結合多種影像前處理技術、深度學習模型優化及後處理技術的 3D 胰臟分割方法。這種方法主要包括對 3D U-Net 模型架構的改進，並加入卷積塊注意力模組 (Convolutional block attention module, CBAM)、深度監督 (Deep Supervision) 和類金字塔池化 (Pyramid-like Pooling) 等技術。此外，藉由使用限制對比度自適應直方圖均衡化來增強影像對比度 (CLAHE)，Coarsed Segmentation Network 進行初步定位，並且以影像平滑化機制及雜訊過濾作為影像邊界修正後處理，以提高胰臟分割準確率。 實驗結果顯示，本研究之方法於相同流程中在NIH TCIA Pancreas-CT 資料集上，並且以交叉驗證方法，取得了比其他 CNN-based 3D 模型更高的分割準確率。;In this study, we investigated a deep learning-based pancreas segmentation method. With the continuous progress of medical imaging technology, the application of pancreas segmentation in medical diagnosis has become increasingly important. Since the pancreas is located deep in the abdominal cavity, its texture features are complex and highly variable, which makes automatic segmentation a huge challenge.
In this study, we proposed a 3D pancreas segmentation method that combines various image pre-processing techniques, deep learning model optimization, and post-processing techniques. This method primarily involves the improvement of the 3D U-Net model architecture and the addition of techniques such as Convolutional block attention module (CBAM), Deep Supervision, and Pyramid-like Pooling. In addition, we enhanced image contrast by using Constrained Contrast Adaptive Histogram Equalization (CLAHE), Coarsed Segmentation Network for the preliminary localization, and an image smoothing mechanism and noise filtering are used as post-processing of image boundary correction to improve the accuracy of pancreas segmentation.
The experimental results show that the method in this study achieves higher segmentation accuracy than other CNN-based 3D models in the same process on the NIH TCIA Pancreas-CT dataset with the K-Fold cross-validation method.
