[Paper Review] Lung Nodules Detection and Segmentation Using 3D Mask-RCNN

Evi Kopelowitz, Guy Engelhard|arXiv (Cornell University)|Jul 17, 2019

Lung Cancer Diagnosis and Treatment13 references38 citations

TL;DR

The paper adapts MaskRCNN to 3D CT scans to detect and segment lung nodules end-to-end, achieving competitive detection and segmentation on LUNA16.

ABSTRACT

Accurate assessment of Lung nodules is a time consuming and error prone ingredient of the radiologist interpretation work. Automating 3D volume detection and segmentation can improve workflow as well as patient care. Previous works have focused either on detecting lung nodules from a full CT scan or on segmenting them from a small ROI. We adapt the state of the art architecture for 2D object detection and segmentation, MaskRCNN, to handle 3D images and employ it to detect and segment lung nodules from CT scans. We report on competitive results for the lung nodule detection on LUNA16 data set. The added value of our method is that in addition to lung nodule detection, our framework produces 3D segmentations of the detected nodules.

Motivation & Objective

• Motivate automated, end-to-end 3D detection and segmentation of lung nodules to improve radiologist workflow and measurement accuracy.
• Extend MaskRCNN to 3D medical images and small object detection.
• Demonstrate the model's ability to provide 3D nodule segmentations along with detections.
• Evaluate detection performance against existing methods and radiologist benchmarks on LUNA16.
• Assess segmentation accuracy and the correlation between predicted volumes and ground truth.

Proposed method

• Adapt 2D MaskRCNN architecture to 3D CT data with a 128^3 patch input and 0.5 mm voxel spacing.
• Use a 3D Inception-ResNet-v2 backbone with atrous convolutions to preserve resolution.
• Employ 3D RPN and ROI Align to propose and refine nodules, with a MASK head for 3D segmentation.
• Train end-to-end with focal loss for classification and IOU loss for MASK segmentation heads.
• Perform 10-fold cross-validation and heavy data augmentation to mitigate overfitting.
• Apply overlapping window inference with NMS and lung-mask filtering to reduce false positives.

Experimental results

Research questions

• RQ1 Can a single end-to-end 3D network detect and segment lung nodules from CT scans effectively?
• RQ2 What are the detection metrics (e.g., CPM, sensitivity, FP rate) achieved by 3DMaskRCNN on LUNA16?
• RQ3 How does 3DMaskRCNN's segmentation performance (DSC) compare to radiologists and other methods?
• RQ4 Is there a strong correlation between predicted nodule volumes and ground-truth volumes?
• RQ5 What is the impact of post-processing steps (FP reduction, lung masking) on performance?

Key findings

• 3DMaskRCNN achieves CPM of 0.86 with FP reduction and 0.826 without FP reduction on LUNA16.
• Sensitivity, FP per scan, and CPM are competitive with or exceed several prior methods and radiologist benchmarks.
• DSC for 3DMaskRCNN segmentation is 70±10, which is close to radiologists (76±16) and other methods (CNN on diameter 79±19; PN-SAMP-S1 74±3.57).
• The model shows a strong correlation (0.96) between predicted nodule volumes and ground-truth volumes.
• Hausdorff distance for boundaries is 2.49 mm ± 2.05 mm, suggesting boundary precision is limited by small object size.
• 3DMaskRCNN provides both detection and full 3D segmentation in a single end-to-end framework, with competitive performance on LUNA16.