Tumour segmentation

Pipeline

Whole slides become patches, then a dataset; model predictions are evaluated against reference masks. — Whole-slide images are too large for a model, so they are cut into 256 × 256 patches with matching tumour masks. CNN and ResNet classify each patch as tumour or not; U-Net labels tumour pixels.

Example segmentations

Four tissue patches beside ground-truth masks and two U-Net predictions, showing agreement and differences. — Left to right: tissue, reference mask, and the best U-Nets trained on 20 (Model 1) and 45 (Model 2) whole-slide images.

Open full-size comparison. Swipe horizontally on mobile, or focus the figure and use arrow keys.

Results

U-Net results on slides 46 to 50, by number of training slides and the tumour share of the training data.

Training slides	Training mix	Mean IoU	Accuracy	F1	Recall	Precision
20	50% tumour	40.7%	70.4%	68.7%	60.9%	72.3%
20	67% tumour	40.7%	70.3%	64.2%	55.6%	53.3%
20	100% tumour	18.8%	37.7%	29.2%	87.5%	37.7%
45	50% tumour	24.2%	48.5%	41.0%	35.5%	100.0%
45	67% tumour	46.9%	82.8%	85.7%	84.5%	41.1%
45	100% tumour	15.7%	31.5%	27.9%	100.0%	31.5%

Balance decides the result. Training only on tumour data overfits to tumour, scoring zero on benign tissue. With 45 slides, a 50/50 mix swings the other way, because the larger set holds many more benign patches. The strongest result came from 45 slides with two-thirds tumour data.

Muhanad Tuameh · Emre Arslanoğlu

Models and notebooks

PyTorch CNN, ImageNet-pretrained ResNet18, and TensorFlow/Keras U-Net.

Project README · Notebooks · Report