Transformer-Based CT Anomaly Detection and Auto-Segmentation of Sparse Lung Nodules
Keywords: Transformer, CT scans, lung nodules, anomaly detection, auto-segmentation, Deformable-DETR, sparse data, medical imaging, self-attention, multi-scale learning, object detection, Focal Loss, segmentation
TL;DR: A Transformer-based model that combines anomaly detection and auto-segmentation to accurately identify and segment sparse lung nodules in CT scans.
Abstract: Accurate segmentation of lung nodules in computed tomography (CT) scans is challenging due to extreme class imbalance, where nodules appear sparsely among healthy tissue. Lung tumor boards often review these scans manually, a time-consuming process. This paper introduces a novel two-stage approach for lung tumor segmentation by framing the problem as anomaly detection. The method is divided into two stages, allowing each model to leverage its strengths. Stage 1 focuses on region proposal, employing a custom Deformable Detection Transformer with Focal Loss to overcome class imbalance and localize sparse tumors. In Stage 2, the predicted bounding boxes are refined into pixel-wise segmentation masks using a fine-tuned variant of Meta's Segment Anything Model (SAM) for semantic segmentation. To address the challenge of nodule sparsity and improve spatial context, a 7.5 mm Maximum Intensity Projection (MIP) is applied, aiding in the differentiation between nodules, bronchioles, and vascular structures. The model achieves a Dice coefficient of 92.4%, with 95.2% sensitivity and 93.2% precision on the LUNA16 dataset, demonstrating robust performance in real-world clinical conditions where nodule sparsity is 5%.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12414
Loading