MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

CT Radiomics Feature Extraction and Machine Learning to Stratify Cellular Subtypes of Benign Pulmonary Nodules

S Tu1*, (1) Chang Gung University, Tao-Yuan,Taiwan, (2) Linkuo Chang Gung Memorial Hospital, Tao-Yuan, Taiwan

Presentations

(Sunday, 7/29/2018) 3:30 PM - 4:00 PM

Room: Exhibit Hall | Forum 6

Purpose: In pulmonary lung cancer diagnosis, the reactive lymphoid hyperplasia tissue and benign parenchymal lesion are classified as a benign pulmonary nodule. However a benign parenchymal lesion may gradually evolve its malignancy risk as a malignant tumor. Therefore, it is important to differentiate these two types of benign nodules at the early stage. We used radiomics features extracted from thin-section CT images and machine learning which was supervised by the pathological examination to differentiate between these lymphoid hyperplasia tissues and benign parenchymal lesions.

Methods: This work was retrospective and approved by our institutional review board. 50 image sets of thin-section CT and their reports of pathological diagnosis were reviewed. These images were acquired from the GE BrightSpeed 16 scanner. The slice-thickness of these thin-section CT images was 0.625 mm. The software program of Imaging Biomarker Explorer (IBEX) was used in the nodule delineation and radiomics feature extraction. These nodules were delineated by an experienced radiologist. 275 image features were extracted from the IBEX program. The multiple-comparison test and a p-value threshold of 005 were used to evaluate the differentiation performance of these radiomics features. An ensemble machine learning of random forest was used to classify these two types of benign nodules. We used the 10-fold cross-validation and receiver operating characteristic curve (ROC) to evaluate the performance of machine learning model.

Results: 28 out of 275 features were found useful for differentiation between two subtypes of benign nodules. The accuracy of machine learning classification was 81% with the 10-fold cross-validation. The area under the ROC curve was 0.82.

Conclusion: The technical approach of CT radiomics feature extraction and machine learning which was supervised by the pathological diagnosis were demonstrated useful help the radiologist to stratify two subtypes of benign pulmonary nodules between the reactive lymphoid hyperplasia tissue and benign parenchymal lesion.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email