Automatic lung nodule detection in thoracic CT scans using dilated slice-wise convolutions.

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computer-aided detection (CADe) systems

Background:

• Current automated medical image analysis for volumetric data often uses 2D or 3D convolutional neural networks (CNNs).
• These methods may not fully leverage the contextual information present in volumetric medical data.
• There is a need for integrated approaches that combine the strengths of both 2D and 3D CNNs.

Purpose of the Study:

• To develop a novel unified CNN-based model for analyzing volumetric medical images.
• To combine the advantages of 2D and 3D CNNs in a single framework.
• To improve the performance of medical image analysis tasks through a hybrid approach.

Main Methods:

• A unified CNN model extracting multiscale contextual information from 2D slices.
• Utilizing dilated 1D convolutions across slices to aggregate in-plane features.
• Implementing a curriculum learning strategy for a two-stage screening and false positive reduction system.

Main Results:

• The proposed approach was evaluated using a computer-aided detection (CADe) system for lung nodules.
• Achieved a sensitivity of > 0.99 in the screening stage and > 0.96 at 8 false positives per case in the reduction stage.
• Demonstrated effective analysis of volumetric data on 888 CT exams.

Conclusions:

• The novel unified CNN model provides competitive results compared to existing 3D frameworks.
• Curriculum learning strategies significantly benefit two-stage systems commonly used in medical imaging.
• The proposed method offers an effective approach for volumetric medical image analysis.