RenseNet: A Deep Learning Network Incorporating Residual and Dense Blocks with Edge Conservative Module to Improve...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer-Aided Diagnosis

Background:

• Deep learning excels in medical image analysis, with target classification and model interpretability being key applications.
• Existing deep learning algorithms often use pooling operations for feature extraction, which can degrade image details crucial for detecting small objects.
• This degradation poses challenges in accurately identifying small lesions in medical imaging.

Purpose of the Study:

• To design and validate a novel deep learning architecture, RenseNet, for improved medical image analysis.
• To address the limitations of pooling operations in preserving fine image details, particularly for small lesions.
• To enhance both the inference and interpretation capabilities of deep learning models in medical imaging.

Main Methods:

• Developed a Rense block, integrating skip connections from residual and dense blocks, with mathematical validation.
• Introduced an edge-conservative module featuring a compensation path to prevent feature blurring during pooling.
• Evaluated the RenseNet architecture on two independent CT datasets: kidney stones and lung tumors.

Main Results:

• RenseNet demonstrated superior classification performance and model interpretability compared to state-of-the-art methods.
• The model effectively preserved image details, crucial for identifying small lesions often missed by conventional approaches.
• Explanation heatmaps from RenseNet provided clear insights into the model's decision-making process.

Conclusions:

• The proposed RenseNet effectively mitigates information loss associated with pooling operations in deep learning.
• RenseNet shows significant potential for improving the diagnosis and treatment of diseases characterized by small lesions.
• This advancement contributes to more efficient and accurate medical image analysis, particularly for challenging cases.