Learning from Experts: Developing Transferable Deep Features for Patient-Level Lung Cancer Prediction

Due to recent progress in Convolutional Neural Networks (CNNs), developing image-based CNN models for predictive diagnosis is gaining enormous interest. However, to date, insufficient imaging samples with truly pathological-proven labels impede the evaluation of CNN models at scale. In this paper, we formulate a domain-adaptation framework that learns transferable deep features for patient-level lung cancer malignancy prediction. The presented work learns CNN-based features from a large discovery set (2272 lung nodules) with malignancy likelihood labels involving multiple radiologists’ assessments, and then tests the transferable predictability of these CNN-based features on a diagnosis-definite set (115 cases) with true pathologically-proven lung cancer labels. We evaluate our approach on the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset, where both human expert labeling information on cancer malignancy likelihood and a set of pathologically-proven malignancy labels were provided. Experimental results demonstrate the superior predictive performance of the transferable deep features on predicting true patient-level lung cancer malignancy (Acc = 70.69 %, AUC = 0.66), which outperforms a nodule-level CNN model (Acc = 65.38 %, AUC = 0.63) and is even comparable to that of using the radiologists’ knowledge (Acc = 72.41 %, AUC = 0.76). The proposed model can largely reduce the demand for pathologically-proven data, holding promise to empower cancer diagnosis by leveraging multi-source CT imaging datasets.