Wednesday 12 March 2025
Deep learning algorithms have long been touted as a solution for automatically identifying blood vessels in histology images, crucial for diagnosing and monitoring cancer. However, these models often struggle to accurately segment the tiny vessels due to their varied appearances and limited training data.
A team of researchers has now developed a novel approach that addresses this challenge by generating high-contrast guiding maps from the input images. These maps highlight the relevant features of blood vessels, allowing deep learning models to focus on what matters most.
The team used a dataset of 28 regions of interest from H&E-stained oesophageal tissue samples, with each region containing both normal and tumour blood vessels. They generated guiding maps by applying a series of colour space conversions, thresholding, and morphological operations to the input images.
These maps were then concatenated with the original RGB image to create a 4-channel input for training deep learning models. The researchers compared the performance of various models, including U-Net, FCN, TransUnet, EfficientNet B0, B2, B6, and V2-M, using both RGB and RGB+guiding map inputs.
The results showed that incorporating guiding maps significantly improved segmentation accuracy across all models. For example, the EfficientNet B2 model achieved a 5% increase in DSC (Dice Similarity Coefficient) when using guiding maps for segmenting tumour blood vessels. The improvement was even more pronounced for larger models like EfficientNet V2-M.
The study’s findings suggest that generating high-contrast guiding maps can be an effective way to enhance the performance of deep learning algorithms for blood vessel segmentation in histology images. This approach could have significant implications for cancer diagnosis and monitoring, where accurate identification of blood vessels is critical for understanding tumour biology and developing targeted therapies.
The researchers hope to expand their dataset to include a wider range of tissue types and investigate the relationship between blood vessel density and spatial distribution with tumour metabolism and prognosis. If successful, this could lead to more effective personalised treatment strategies for cancer patients.
By leveraging the strengths of deep learning algorithms and cleverly generated guiding maps, scientists may be able to unlock new insights into the complex interactions between tumours and their microenvironments. With potential applications in both research and clinical practice, this work has the potential to revolutionise our understanding of cancer biology and improve patient outcomes.
Cite this article: “Guiding Maps Enhance Blood Vessel Segmentation in Histology Images for Cancer Diagnosis”, The Science Archive, 2025.
Histology Images, Blood Vessels, Deep Learning, Image Segmentation, Cancer Diagnosis, Guiding Maps, Colour Space Conversions, Thresholding, Morphological Operations, Tumour Biology
