Deep multi-task learning for nephropathy diagnosis on immunofluorescence images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nephrology

Background:

• Immunofluorescence (IF) imaging is crucial for nephropathy diagnosis but often suffers from blur, hindering deep neural network (DNN) performance.
• Existing DNNs for nephropathy diagnosis show reduced accuracy on blurred IF images, limiting real-world application.

Purpose of the Study:

• To develop a robust method for nephropathy diagnosis using blurred immunofluorescence images.
• To bridge the gap between low-level image processing and high-level medical diagnostic tasks.

Main Methods:

• Established two IF image databases (IFVB, Real-IF) with synthetic and real-world blurs.
• Proposed a Deep Hierarchical Multi-Task Learning based Nephropathy Diagnosis (DeepMT-ND) method.
• DeepMT-ND simultaneously performs nephropathy diagnosis, image quality assessment (IQA), and de-blurring.

Main Results:

• DeepMT-ND demonstrated superior diagnostic accuracy over human nephrologists, with improvements of 15.4% (IFVB) and 6.5% (Real-IF).
• The method achieved comparable performance in auxiliary IQA and de-blurring tasks.
• Validated effectiveness on both synthetic and real-world blurred IF image datasets.

Conclusions:

• DeepMT-ND offers a novel hierarchical multi-task learning framework for improved nephropathy diagnosis on blurred IF images.
• The method enhances the potential of DNNs in clinical nephropathy diagnosis scenarios.
• Experimental results confirm the diagnostic accuracy and generalization ability of DeepMT-ND.