Deep learning for lung cancer prognostication: A retrospective multi-cohort radiomics study

• DOI: 10.1371/journal.pmed.1002711
• Publication Date: 2018-11-30T18:24:19Z
• Authors: Ahmed Hosny; Chintan Parmar; Thibaud P. Coroller; Patrick Grossmann; Roman Zeleznik; Avnish Kumar; Johan Bussink; Robert J. Gillies; Raymond H. Mak; Hugo J. W. L. Aerts
• Keywords: Adult; Male; Lung Neoplasms; Clinical Decision-Making; Risk Assessment; 03 medical and health sciences; Deep Learning; 0302 clinical medicine; Predictive Value of Tests; Risk Factors; Carcinoma, Non-Small-Cell Lung; Humans; Diagnosis, Computer-Assisted; Aged; Neoplasm Staging; Retrospective Studies; Aged, 80 and over; R; Reproducibility of Results; Radboudumc 9: Rare cancers RIHS: Radboud Institute for Health Sciences; Middle Aged; 3. Good health; Medicine; Radiographic Image Interpretation, Computer-Assisted; Female; Research Article; Preliminary Data

Non-small-cell lung cancer (NSCLC) patients often demonstrate varying clinical courses and outcomes, even within the same tumor stage. This study explores deep learning applications in medical imaging allowing for automated quantification of radiographic characteristics potentially improving patient stratification.We performed an integrative analysis on 7 independent datasets across 5 institutions totaling 1,194 NSCLC (age median = 68.3 years [range 32.5-93.3], survival 1.7 0.0-11.7]). Using external validation computed tomography (CT) data, we identified prognostic signatures using a 3D convolutional neural network (CNN) treated with radiotherapy (n 771, age 68.0 1.3 We then employed transfer approach to achieve surgery 391, 69.1 37.2-88.0], 3.1 0.0-8.8]). found that CNN predictions were significantly associated 2-year overall from start respective treatment (area under receiver operating characteristic curve [AUC] 0.70 [95% CI 0.63-0.78], p < 0.001) (AUC 0.71 0.60-0.82], patients. The was also able stratify into low high mortality risk groups both (p 0.03) datasets. Additionally, outperform random forest models built parameters-including age, sex, node metastasis stage-as well as robustness against test-retest (intraclass correlation coefficient 0.91) inter-reader (Spearman's rank-order 0.88) variations. To gain better understanding captured by CNN, regions most contribution towards highlighted importance tumor-surrounding tissue stratification. present preliminary findings biological basis phenotypes being linked cell cycle transcriptional processes. Limitations include retrospective nature this opaque black box networks.Our results provide evidence networks may be used stratification based standard-of-care CT images motivates future research deciphering prospective data.