Edwin Wang

PhD

University of Calgary

Professor

NLST (Learn more about this study)

NLST-340

Aug 16, 2017

Predict histological grade by CT image auto-reviewing

Low-dose computed tomography (LDCT) is the recommended screening test for lung cancer. LDCT scans the chest and the reviewers evaluate the generated images for nodules or abnormalities present. This process is mainly done manually by radiologists, and has a relatively high false-positive rate along with over-diagnoses. As true-positive LDCT image data in NLST is paired with pathological images, tissue microarrays and clinical descriptions, we’re willing to design a texture analysis method based on machine learning to predict histological grades and improve the diagnosis.

To reach the overall goal, we have several specific aims:

(1) To generate pix2pix models with bounding box to automatically detect nodules in LDCT images. (eg, to find region of interest(ROI) with segmentation)
(2) To train & evaluate classification models to predict histological grades and survival length of ROIs.
(3) To use the feature extraction headers learned in above models, build natural language generation headers to give rough descriptions of LDCT images.