High performance with fewer labels using semi-weakly supervised learning for pulmonary embolism diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate pulmonary embolism (PE) detection on CT pulmonary angiography (CTPA) is crucial for patient care.
• Exhaustive annotation of medical images for supervised learning is resource-intensive and time-consuming.

Purpose of the Study:

• To develop and evaluate a semi-weakly supervised learning approach for PE detection on CTPA.
• To reduce the burden of medical image annotation without compromising diagnostic performance.

Main Methods:

• Attention-based Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) models were trained on the RSNA PE CT dataset.
• Three training configurations were used: weak (examination-level labels), strong (all labels), and semi-weak (examination-level plus limited slice-level labels).
• Models were externally validated on pooled Aida and FUMPE datasets.

Main Results:

• Semi-weakly supervised models achieved an Area Under the Curve (AUC) of 0.928 using only ~25% of slice-level labels, closely matching the strongly supervised model's AUC of 0.932.
• External validation demonstrated high performance, with AUCs of 0.999 for semi-weak and 1.000 for strong models.
• The proposed method significantly reduces labeling requirements while maintaining diagnostic accuracy.

Conclusions:

• Semi-weakly supervised learning offers an effective strategy for PE detection on CTPA, mitigating annotation burdens.
• This approach accelerates the development and clinical integration of AI models for medical imaging.
• The findings suggest a pathway to enhance patient care through more efficient AI model deployment.