Improving Lung Cancer Diagnosis and Survival Prediction with Deep Learning and CT Imaging

TL;DR

This work addresses improving lung cancer diagnosis and survival prediction from a single CT scan by linking image-derived morphology to hazard via an extended Cox model replaced by a 3D CNN. A scalable mini-batched loss approximates the Cox partial likelihood, enabling large-scale training, and a two-task objective jointly optimizes cancer classification (AUC) and survival risk (C-index). The approach is validated through simulations on MNIST and Nodule-CIFAR and demonstrated on NLST CT data with multiple 3D architectures, showing superior or competitive performance relative to baselines. The method promises practical clinical impact by enabling simultaneous screening and prognosis from one image, potentially guiding early detection and treatment decisions across 3D medical imaging domains.

Abstract

Lung cancer is a major cause of cancer-related deaths, and early diagnosis and treatment are crucial for improving patients' survival outcomes. In this paper, we propose to employ convolutional neural networks to model the non-linear relationship between the risk of lung cancer and the lungs' morphology revealed in the CT images. We apply a mini-batched loss that extends the Cox proportional hazards model to handle the non-convexity induced by neural networks, which also enables the training of large data sets. Additionally, we propose to combine mini-batched loss and binary cross-entropy to predict both lung cancer occurrence and the risk of mortality. Simulation results demonstrate the effectiveness of both the mini-batched loss with and without the censoring mechanism, as well as its combination with binary cross-entropy. We evaluate our approach on the National Lung Screening Trial data set with several 3D convolutional neural network architectures, achieving high AUC and C-index scores for lung cancer classification and survival prediction. These results, obtained from simulations and real data experiments, highlight the potential of our approach to improving the diagnosis and treatment of lung cancer.

Figures (10)

• Figure 1: Two-task Convolution Neural Network Illustration.
• Figure 2: Simulated Survival Time Distributions. (a) Survival time distributions for the two digits in Simulation A, without the censoring mechanism; (b) Survival time distributions for the two digits in Simulation B, with the censoring mechanism. The censored cases are labeled in orange, which overlaps the upper half of the event cases.
• Figure 3: Simulation Losses by Epoch. (a) Simulation A (b) Simulation B.
• Figure 4: A Nodule-CIFAR Example. Non-cancer cases only have small black and white dots scattered over the images, simulating benign nodules. In addition to benign nodules, cancer cases have 2 larger white patches to simulate malignant nodules.
• Figure 5: Nodule Size and Survival Time Distribution by Group. (a) Nodule size distribution by group. The non-cancer group has smaller nodules on average when compared with the cancer group. Within the cancer group, event cases (those who eventually die of cancer in simulation) have larger malignant nodules. (b) Survival time distribution by group in Nodule-CIFAR. The time-to-event for the non-cancer group is larger than the cancer group. Within the cancer group, the time-to-event of censored is larger than that of the event cases.