Prediction of adenocarcinoma among other subtypes of lung cancer from CT using deep learning.

Abstract

3057 Background: Determination of histological subtype is a crucial step in the management of patients with lung cancer as it informs prognosis and management. The identification of adenocarcinoma (AC) is particularly important with new targeted treatments becoming available. Although the gold standard for diagnosing histological subtype is pathological analysis of tissue samples, interventions can present a risk of complication. Imaging-based, computational approaches to distinguishing malignant from benign lesions have shown promising results. A similar approach may also be applied to determining histological subtype, which could provide an early, non-invasive alternative or complimentary method to biopsy. Here, we investigate an imaging and machine learning method to predict the subtype of a malignant lung lesion. Methods: A dataset of 1493 primary lung cancer patients was collected, of which 943 were diagnosed with AC. The histological subtypes of the non-AC patients were, squamous-cell carcinoma (158), large-cell carcinoma (69), small-cell carcinoma (33), other subtypes (27), or unreported non-AC subtype (253). This consists of retrospectively collected CT images and demographic data from both screening and clinical settings, across 41 academic and community centres from the USA (35 centres), and Europe (6 centres). All patients included were aged ≥ 18 with no history of cancer in the last 5 years. Each CT was manually curated. Given a CT-image of a lung nodule, a Convolutional Neural Network (CNN) was trained to classify nodules as AC or non-AC, using 8-fold cross-validation. A logistic-regression model based on clinical parameters was also trained using the same data and cross-validation. Classification performance was evaluated using the Area-Under-the-ROC-Curve (AUC), sensitivity, and specificity. Confidence intervals and P values were calculated by nonparametric bootstrapping. Results: The median age of AC patients was 66 yr (non-AC 67 yr) and 62.3% of them were male (non-AC 57.8%). The median pack years for AC patients was 38.4 p.yr (non-AC 50.0 p.yr). For AC tumours, the median diameter was 14.0 mm (non-AC 14.0 mm) and the mean diameter was 15.0 mm (non-AC 16.4 mm). The AC-classification results are tabulated below. The CNN classification performance is significantly better than the logistic-regression baseline across all three measures of performance P <.001. AC classification results. All values are given in % with 95% confidence interval bounds in parentheses. Conclusions: We find that the CNN significantly outperforms a logistic model in identifying AC from other histological subtypes. With further development, this algorithm could prove a useful tool to aid management of lung cancer patients.[Table: see text]