Project 04 / 04Computer VisionHistopathology · Deep Learning

Colorectal Cancer Tissue Classification (Deep Learning)

A custom CNN that beat VGG16 transfer learning at classifying eight histopathology tissue types.

04 · Computer VisionOverview

The build

How it came together.

01
The problem
Colorectal cancer diagnosis leans heavily on manual histopathology, which is slow, expertise-driven, and prone to variability between readers. I wanted to see how far a from-scratch deep-learning model could go at automating tissue classification, working with the TensorFlow colorectal_histology dataset: 5,000 images at 150×150 resolution across 8 tissue types (tumour, debris, stroma, lymphocytes, complex, mucosa, adipose, empty).
02
The model
I designed a custom CNN with convolutional blocks scaling from 32 to 256 filters, using batch normalisation, max-pooling, and dropout to keep it stable and regularised. On held-out data it reached 0.94 accuracy and 0.94 macro-F1, which was strong and, importantly, even across all eight classes rather than carried by the easy ones.
03
The benchmark
To know whether that was actually good, I benchmarked it against the standard move: VGG16 transfer learning with ImageNet weights and fine-tuning. That reached 0.89 accuracy and 0.89 macro-F1, respectable but below my custom network. The takeaway I drew is the interesting part: on this task, domain-specific feature learning outperformed generic pretrained features. Transfer learning isn't automatically the right answer.

Visuals

03 figures

Fig. 02
Training curves: custom CNN (Model 1) against VGG16 transfer learning (Model 2).
Fig. 03
VGG16 confusion matrix: 0.89 accuracy, trailing the custom CNN's 0.94.