Lung Tumor Subtyping and Nucleus Segmentation

• Deep-learning based classification pipeline for subtyping lung tumors from histology.
• Assessing the impact of nucleus segmentation on tumor discernibility.

Link to the Research Paper [HTML].

Cite us

If you find our work useful in your research, please cite us!

@article{jsm2023,
  title={A deep learning approach for nucleus segmentation and tumor classification from lung histopathological images},
  author={Jaisakthi, SM and Desingu, Karthik and Mirunalini, P and Pavya, S and Priyadharshini, N},
  journal={Network Modeling Analysis in Health Informatics and Bioinformatics},
  volume={12},
  number={1},
  pages={22},
  year={2023},
  publisher={Springer},
  url={https://link.springer.com/article/10.1007/s13721-023-00417-2},
  doi={https://doi.org/10.1007/s13721-023-00417-2}
}

Study Overview

The study method is summarized as a brief algorithm below.

Study Dataset and Data Processing

The histopathology image dataset is sourced from LC25000 Dataset.

• 768 x 768 resolution images of lung histology.
• Contains patch-level labels of tumor type.
• HIPAA compliant and validated source.
• Detailed data description.

Semi-automated nuclear region annotation

1. Automated nuclear region annotation were obtained using a stain-based color thresholding approach, detailed in the algorithm below.
   
2. The obtained annotations were corrected and validated by expert pathologists.
3. Multiple pathologist corrections were compared and averaged. The inter-rater agreement was assessed using generalized conformity index (GCI); a GCI of 0.89 was obtained.

Common Downstream Tumor Classification

• The classifier is a custom lightweight Convolution Neural Network that performs downstream tumor subtyping.

• The common downstream serves the role of a discriminator reference to compare subtyping performances with and without prior nucleus segmentation of histology images.

  

Intermediate Nucleus Segmentation in the P_seg Pipeline

• Rationale: The nuclei portray sufficiently distinct visual characteristics under each tumor type to discern them apart.
• Nuclear regions of the lung histology images are segmented out before classification.
• A segmentation architecture derived from the Xception-style UNet is trained and fine-tuned to automate this nucleus segmentation.