研究目的
To develop an innovative framework for pixel-level semantic segmentation of the tumour border area in colorectal liver metastasis (CRLM) histopathology images by integrating features from deep convolutional networks with spatial and statistical information of cells.
研究成果
The proposed framework successfully integrates multi-dimensional features from deep convolutional networks and cell properties for robust semantic segmentation of tumour border areas in CRLM histopathology images. It demonstrates significant improvement over conventional methods, with high accuracy metrics. However, limitations in nuclei detection highlight the need for further improvements in supervised learning approaches.
研究不足
The performance is partially dependent on the accuracy of nuclei localization and segmentation, which is affected by overlapping and high shape variation in tumour cells. The cell-level model was trained in a weakly supervised manner due to lack of cell-level ground truth, reducing overall performance. Future work should focus on supervised models for nuclei detection and segmentation.
1：Experimental Design and Method Selection:
The framework integrates a two-level deep neural network (cell-level and tissue-level models) with a region-growing algorithm for segmentation. Methods include colour deconvolution, nuclei detection using adaptive thresholding and active contour models, and training of deep convolutional neural networks (AlexNet for tissue level and a custom FCN-5 for cell level).
2：Sample Selection and Data Sources:
Data comes from whole slide images (WSIs) of CRLM patients, with 18 WSIs annotated by pathologists. Patches of size 224x224 for tissue level and 128x128 for cell level are cropped from 40X magnification images.
3：List of Experimental Equipment and Materials:
Titan X Pascal GPU (donated by NVIDIA Corporation) for training deep learning models. Software includes implementations of Non-negative Matrix Factorization (NMF), total variation denoising (TVD), adaptive thresholding, morphological snake model, and deep neural networks.
4：Experimental Procedures and Operational Workflow:
Pre-processing involves colour deconvolution to separate H and E channels, nuclei detection using adaptive thresholding and active contour refinement. Training of DCNNs with stochastic gradient descent (SGD) and data augmentation (geographic transforms and colour perturbation). Segmentation is finalized using a tissue region-growing algorithm based on nuclei properties.
5：Data Analysis Methods:
Evaluation metrics include Pixel Accuracy (PA), F1-score, and Mean Intersection over Union (MIoU) for segmentation performance on test datasets.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
