Project
CRCCD: Colorectal Cancer CNN Benchmark
A deep learning benchmark evaluating 10 convolutional neural network architectures for histopathological colorectal cancer detection, achieving 93.9% accuracy with DenseNet169.
This study benchmarks deep learning architectures on the CRCCD_V1 histopathology image dataset to evaluate performance trade-offs in automated colorectal cancer tissue classification.
The benchmark compares ten distinct convolutional neural network configurations, including transfer learning backbones (DenseNet169, DenseNet121, InceptionResNetV2, ResNet50), edge-oriented models (MobileNetV2, NASNetMobile), and a custom convolutional network (OzNET) trained from scratch. DenseNet169 produced the leading result, reaching 93.91% test accuracy, a 0.9388 weighted F1-score, and a Cohen’s Kappa of 0.9344.
The full pipeline covers data normalization, augmentation, learning rate scheduling, confusion matrices, and ROC-AUC curves, packaged into a single reproducible notebook designed for execution on GPU-accelerated environments. For experiment notebooks and comprehensive metric breakdowns, read more in the GitHub repository.