Artificial Intelligence-Guided Segmentation and Path Planning Software for Transthoracic Lung Biopsy
Purpose: To validate the sensitivity and specificity of a 3 -dimensional (3D) convolutional neural network (CNN) artificial intelligence (AI) software for lung lesion detection and to establish concordance between AI -generated needle paths and those used in actual biopsy procedures. Materials and M...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Published: |
Elsevier
2024
|
| Subjects: | |
| Online Access: | http://eprints.um.edu.my/45279/ https://doi.org/10.1016/j.jvir.2024.02.006 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Purpose: To validate the sensitivity and specificity of a 3 -dimensional (3D) convolutional neural network (CNN) artificial intelligence (AI) software for lung lesion detection and to establish concordance between AI -generated needle paths and those used in actual biopsy procedures. Materials and Methods: This was a retrospective study using computed tomography (CT) scans from 3 hospitals. Inclusion criteria were scans with 1-5 nodules of diameter >= 5 mm; exclusion criteria were poor -quality scans or those with nodules measuring <5mm in diameter. In the lesion detection phase, 2,147 nodules from 219 scans were used to develop and train the deep learning 3D -CNN to detect lesions. The 3D -CNN was validated with 235 scans (354 lesions) for sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) analysis. In the path planning phase, Bayesian optimization was used to propose possible needle trajectories for lesion biopsy while avoiding vital structures. Software -proposed needle trajectories were compared with actual biopsy path trajectories from intraprocedural CT scans in 150 patients, with a match defined as an angular deviation of <5 (degrees) between the 2 trajectories. Results: The model achieved an overall AUC of 97.4% (95% CI, 96.3%-98.2%) for lesion detection, with mean sensitivity of 93.5% and mean specificity of 93.2%. Among the software -proposed needle trajectories, 85.3% were feasible, with 82% matching actual paths and similar performance between supine and prone/oblique patient orientations ( P = .311). The mean angular deviation between matching trajectories was 2.30 degrees (SD +/- 1.22); the mean path deviation was 2.94 mm (SD +/- 1.60). Conclusions: Segmentation, lesion detection, and path planning for CT -guided lung biopsy using an AI -guided software showed promising results. Future integration with automated robotic systems may pave the way toward fully automated biopsy procedures. |
|---|