Deep learning for colon cancer histopathological images analysis

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• Digital pathology utilizes Whole Slide Images (WSIs) for cancer diagnosis and prognosis.
• High resolution and size of WSIs, coupled with sparse annotations, pose challenges for existing diagnostic methods.
• Deep Learning (DL) offers a promising solution for analyzing large-scale histopathological image data.

Purpose of the Study:

• To employ Deep Learning (DL) architectures for classifying and segmenting colon cancer regions within histopathological images.
• To address the limitations of sparsely annotated datasets in digital pathology.
• To identify optimal DL models and training strategies for colon tumor analysis.

Main Methods:

• Reviewed and compared state-of-the-art Convolutional Neural Networks (CNNs): AlexNet, VGG, ResNet, DenseNet, and Inception.
• Utilized transfer learning with the ImageNet dataset to overcome WSI dataset annotation scarcity.
• Implemented pixel-wise segmentation using UNet and SegNet models with a multi-step training strategy, including data augmentation.

Main Results:

• ResNet achieved up to 96.98% accuracy for patch-level classification on the AiCOLO dataset and high performance on public datasets (e.g., 99.98% on a merged dataset).
• SegNet demonstrated strong segmentation performance, reaching up to 81.22% accuracy on the internal dataset and excelling on public datasets (e.g., 99.12% on nct-crc-he-100k).
• The multi-step training strategy proved effective for handling sparse annotations in histopathological images.

Conclusions:

• Deep learning models, especially ResNet for classification and SegNet for segmentation, are highly effective for colon cancer analysis in digital pathology.
• Transfer learning and multi-step training strategies significantly enhance model performance with limited annotated WSI data.
• This study provides valuable insights into selecting suitable DL networks and training approaches for colon tumor segmentation and classification.