Predicting NSCLC prognosis by automated pathology

Presented by:

Tsai, Mu-Hung, M.D.Department of Radiation Oncology, National Cheng Kung University Hospital

2016/09/02

Lung Cancer Most prevalent cancer worldwide Histopathological classification

Small cell carcinoma Non-small cell carcinoma

Squamous cell carcinoma Adenocarcinoma

Grade (Gr.1 – Gr.3) Low level of inter-observer agreement

PathologySquamous cell carcinoma Adenocarcinoma

Aim Fully automated microscopic pathology image features

Predict lung cancer survival

Training

Validation

Methods & Materials The Cancer Genome

Atlas (TCGA) Stanford Tissue

Microarray (TMA) Database

Cross-validate

Image: euthman @ Flickr

Results

Identifying Tumor

Adenocarcinoma vs. Normal Lung

SCC vs. Normal Lung

Adenocarcinoma vs. SCC (TCGA)

Adenocarcinoma vs. SCC (TMA)

Results

Predicting Survival

Predicting Adenocarcinoma Prognosis

Stage IB Grade 3 adenocarcinoma

Predicting SCC Prognosis

Discussion Automated workflow to analyze whole slide pathology

Elastic net-Cox proportional hazards models Computationally efficient Handles right-censored survival data

Performance not very sensitive to ML method

Obtain large number of features!