Exploring Deep Learning Models for Small Histopathology Datasets: Segmentation and Classification of Glomerular...

Area of Science:

• Nephrology
• Computational Pathology
• Artificial Intelligence

Background:

• Glomerular crescent lesions signify severe kidney injury and disease progression.
• Automated identification is challenging due to data limitations, class imbalance, and subtle variations.

Purpose of the Study:

• To develop a robust deep learning (DL) framework for segmenting and classifying glomerular crescent lesions in histopathology images.
• To ensure the framework's reliability and interpretability, especially under limited data conditions.

Main Methods:

• Evaluation of baseline DL models for segmentation (DeepLabV3, U-Net, Transformer-based U-Net, FPN) and classification (EfficientNetV2-B0, ResNet-50, DenseNet-121, CTransPath, RetCCL).
• Development of customized models: CrescentSegNet for segmentation and CrescentDenseNet for classification.
• Assessment of interpretability and reliability using Grad-CAM, saliency mapping, uncertainty estimation, calibration analysis, and t-SNE.

Main Results:

• The proposed CrescentSegNet and CrescentDenseNet models achieved competitive performance on the ISICDM2024 Challenge dataset.
• Cross-dataset evaluations on SICAPv2 and BreaKHis confirmed strong generalization and robustness.
• The framework demonstrated efficiency and interpretability.

Conclusions:

• The developed deep learning framework effectively addresses the challenges in automated identification of glomerular crescent lesions.
• The customized models offer a robust, efficient, and interpretable solution for kidney injury assessment.