Efficient colorectal cancer classification from histopathological images utilizing lightweight convolutional neural...

Area of Science:

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

Background:

• Colorectal cancer is a leading cause of cancer death globally.
• Accurate histopathological analysis is crucial for timely diagnosis and treatment decisions.
• Automated image analysis using deep learning offers potential for improved efficiency and accuracy.

Purpose of the Study:

• To evaluate lightweight convolutional neural network (CNN) variants for binary classification of colon histopathology images.
• To compare the accuracy-efficiency trade-offs of different CNN architectures.
• To assess model robustness against domain shift in colorectal cancer histopathology.

Main Methods:

• A unified framework with standardized preprocessing, data augmentation, and class-weighted optimization was used.
• Four lightweight CNN variants were trained and validated for classifying colon adenocarcinoma and benign tissue.
• Model performance was analyzed using learning curves, confusion matrices, ROC, and precision-recall diagnostics.

Main Results:

• Lite-V2 demonstrated the best accuracy-efficiency trade-off on in-domain validation, with a small footprint and stable convergence.
• Significant performance degradation was observed on independent test sets and external datasets, indicating sensitivity to domain shift.
• Lightweight models showed limitations when faced with distributional variations in histopathological data.

Conclusions:

• Computationally efficient CNNs can aid high-throughput screening and clinical decision support for colorectal cancer.
• Robust clinical deployment necessitates multi-site validation to address data heterogeneity.
• Continued adaptation and development are required for reliable performance across diverse histopathological datasets.