Digital Pathological Image Analysis
Aided by Convolutional Neural Network
Predicts Prognosis of Lung Adenocarcinoma Patients


Lung cancer is a very prevalent cancer type and pathological imaging assessment is an important tool for diagnosis. Previously, pathological imaging reading is done by human doctors. However human evaluation is objective and labor-intensive. In addition, pathological images may contain important prognosis and treatment response information that cannot be reliably understood by humans. Therefore, it is a very important to create machine learning algorithms to automatically recognize pathological imaging data and correlate the automatically retrieved features with patient phenotype.

In this study, we obtained pathological imaging slides and prognosis information of patient from TCGA lung adenocarcinoma project (TCGA dataset), the NLST project (NLST dataset), MDACC Lung Spore project (Spore dataset), and the patients from National Cancer centre/Cancer Hospital of Chinese Academy of Medical Sciences, China (Beijing dataset). We developed an imaging analysis pipeline involving image segmentation and convolutional neural network to recognize the cell types in pathological imaging data. And we showed that the extracted cell type-level features by this pipeline are predictive of lung adenocarcinoma prognosis by training a coxph model. This is an early effort to build image analysis pipelines to recognize cell types on lung cancer pathological images and predict patients’ prognosis based on cell type-level features. This approach can potentially be extended to other types of tumor and may be able to be applied to clinical practices in the future.

KEYWORDS: deep learning, lung adenocarcinoma, pathological imaging, prognosis, tumor heterogeneity


Step 1:  ROI Selection

Step 2:  Cell Segmentation to Cell-Center Extraction

Step 3:  Cell Type Prediction (Convolution Neural Network)

Step 4:   Feature Extraction

Step 5:  Survival Analysis


Download the code for digital pathological image analysis aided by convolutional neural network

Command Line Software

Source codes of the ConvPath software for digital pathological image analysis.

Test Data

Sample test data:

(1) An image file (*.svs file)

(2) A region of interest definition file (*.xml file)

User Manual

User manual explaining installation, usage, and output formats of ConvPath.


Members who contribute to the project

Xie Yang

Xie Yang

Director and Associate Professor
Guanghua Xiao

Guanghua Xiao

Associate Professor
Tao Wang

Tao Wang

Assistant Professor

Shidan Wang

Shidan Wang

PhD Student
Faliu Yi

Faliu Yi

Staff Title




Suite NC8.512, 5323 Harry Hines Blvd., Dallas, TX 75390