MGA-NET: MULTI-SCALE GUIDED ATTENTION MODELS FOR AN AUTOMATED DIAGNOSIS OF IDIOPATHIC PULMONARY FIBROSIS (IPF).

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Idiopathic pulmonary fibrosis (IPF) diagnosis from high-resolution computed tomography (HRCT) scans is challenging, especially with limited labels.
• Weakly supervised learning offers a potential solution for medical image analysis where detailed annotations are scarce.

Purpose of the Study:

• To develop and evaluate a Multi-scale, domain knowledge-Guided Attention network (MGA-Net) for weakly supervised IPF diagnosis.
• To improve the accuracy and explainability of deep learning models for interstitial lung disease (ILD) classification.

Main Methods:

• The MGA-Net model was trained on axial chest HRCT scans from 279 IPF and 423 non-IPF ILD patients.
• Performance was assessed using area under the receiver operating characteristic curve (AUC) with stratified five-fold cross-validation.
• The study compared models with no attention, unguided attention, and multi-scale guided attention modules.

Main Results:

• The MGA-Net model achieved the highest AUC of 0.971 ± 0.021, outperforming models without attention (0.690 ± 0.194) and with unguided attention (0.956 ± 0.040).
• Guided attention maps focused on relevant lung regions, enhancing model interpretability.
• Integrating both high- and medium-resolution guided attention yielded the best diagnostic performance.

Conclusions:

• MGA-Net effectively utilizes domain knowledge for weakly supervised medical image diagnosis.
• The proposed model enhances both the accuracy and explainability of IPF detection in HRCT scans.
• Multi-scale guided attention is crucial for improving diagnostic performance in challenging weakly supervised tasks.