PtbNet: Based on Local Few-Shot Classes and Small Objects to Accurately Detect PTB.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pulmonary Tuberculosis (PTB) is a highly infectious disease requiring early detection for effective management.
• Digital Radiography (DR) is a primary diagnostic tool, but challenges exist in identifying subtle and small lesions.
• Current AI methods struggle with few-shot learning and detecting small objects in PTB diagnosis.

Purpose of the Study:

• To develop an AI model for accurate detection of small and varied Pulmonary Tuberculosis lesions in Digital Radiography.
• To address limitations in few-shot learning and small object detection for PTB diagnosis.

Main Methods:

• Geometric data augmentation and a diffusion probability model were used to enhance the dataset for six few-shot classes.
• A novel multi-lesion detector, PtbNet, based on RetinaNet architecture, was designed for small object detection.
• The model was trained and evaluated on a dataset of Digital Chest Radiographs (DCRs).

Main Results:

• Data augmentation increased the DCR dataset size by 80% (570 to 2,859 images).
• PtbNet achieved an Average Precision (AP) of 28.2, outperforming 9 state-of-the-art methods.
• Ablation studies showed performance gains with integrated components like BiFPN+ and PSPD-Conv.

Conclusions:

• PtbNet significantly improves the detection of small PTB lesions and enhances the uniform detection of diverse PTB types.
• The model aids physicians in accurate PTB lesion diagnosis.
• The developed AI approach offers a promising solution for challenging PTB detection cases.