Deep learning in lung cancer pre- and post-radiation therapy: diagnosis of malignancy and radiation-induced lung injury from 3D x-ray CT.

Lung cancer remains the leading cause of cancer-related mortality worldwide, with early detection and accurate diagnosis being critical for improving patient outcomes. Additionally, the progression of Radiation-Induced Lung Injury (RILI) following Stereotactic Body Radiation Therapy (SBRT) for lung cancer presents a significant diagnostic challenge. This dissertation addresses these challenges by developing deep learning-based diagnostic tools for both pre-treatment lung nodule malignancy classification and post-treatment RILI identification using 3D X-ray CT imaging. The research is divided into two primary objectives. First, for lung nodule malignancy classification, we developed a biopsy-confirmed dataset, called NLSTx, to train and evaluate deep learning models while also leveraging a widely used dataset, LIDC, for validation. We iteratively developed and improved on deep learning techniques for this task, beginning with a novel Convolutional Recurrent Neural Network (CNN-RNN) architecture that was introduced to sequentially process 2D slices of 3D CT volumes, achieving performance comparable to 3D CNNs while requiring fewer parameters. Enhancements were made through attention mechanisms in Siamese networks, resulting in multi-time-point analysis of longitudinal CT data to capture nodule growth, leading to further performance gains. The most significant advancements were then made by using Low-Rank Adaptation (LoRA) fine-tuning of large Vision Transformers (ViTs) pretrained on natural images. LoRA effectively adapts these models to medical imaging tasks, achieving superior ROC AUC scores while significantly reducing the number of trainable parameters and shortening training times. Second, for RILI diagnosis in post-SBRT follow-up CT scans, a 3D CNN was initially trained to identify early radiographic markers of RILI, such as ground-glass opacities and fibrosis. Despite small cohort sizes, the model demonstrated robust performance with an ROC AUC of 0.762, particularly excelling in challenging subsets like scans taken within three months of SBRT and nodules smaller than 2.5 cm. Building upon this, a LoRA-tuned Vision Transformer (DINOv2 ViT) was adapted to further improve efficiency and diagnostic accuracy, achieving competitive performance while reducing computational burden. This dissertation demonstrates significant advancements in deep learning for lung imaging while demonstrating the power of parameter-efficient fine-tuning methods such as LoRA with vision transformers. The findings highlight the potential for real-time clinical integration of automated diagnostic tools to improve lung cancer detection and mitigate post-SBRT complications.