MultiSCCHisto-Net-KD: A deep network for multi-organ explainable squamous cell carcinoma diagnosis with knowledge...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Histopathology image analysis

Background:

• Squamous cell carcinoma (SCC) diagnosis relies on manual analysis of histopathological images, which is time-consuming and subjective.
• Existing deep learning models often focus on high-magnification images and specific organs, neglecting low-magnification structural abnormalities across diverse origins.
• Computational expense of deep neural networks hinders practical application in histopathological image analysis.

Purpose of the Study:

• To develop a robust deep learning model for detecting squamous cell carcinoma (SCC) in low-magnification histopathological images, irrespective of the organ of origin.
• To address the computational cost of deep learning models through knowledge distillation, creating an efficient yet accurate diagnostic tool.
• To enhance model interpretability and build confidence in diagnostic outcomes using explainable AI techniques.

Main Methods:

• Proposed a novel deep neural network, MultiSCCHisto-Net, for detecting SCC in low-magnification histopathological images from various organs.
• Implemented knowledge distillation to train a smaller, efficient student model (MultiSCCHisto-Net-KD) from a complex teacher model, preserving performance and improving generalization.
• Integrated gradient-weighted class activation mapping (Grad-CAM) for explainable AI, visualizing image regions critical for SCC classification.

Main Results:

• The MultiSCCHisto-Net-KD model achieved high accuracy rates of 97% on private multi-centric datasets and 93% on public datasets.
• The model demonstrated superior performance compared to existing state-of-the-art models in SCC detection across diverse histopathological images.
• Explainable AI techniques provided visual evidence supporting the model's classification decisions, enhancing diagnostic confidence.

Conclusions:

• The proposed MultiSCCHisto-Net and its distilled version (MultiSCCHisto-Net-KD) offer a robust, accurate, and computationally efficient solution for SCC detection in low-magnification histopathology.
• The model's ability to generalize across different organs of origin represents a significant advancement in automated computational pathology.
• The integration of knowledge distillation and explainable AI paves the way for more reliable and interpretable AI-driven diagnostic tools in cancer research.