Tumor segmentation in CT

The problem

Outlining a tumor by hand on every slice of a CT scan is slow, and the thing you are looking for is tiny. In a lung scan most voxels are not tumor, so a model can reach a very low loss by predicting “nothing” everywhere. Getting a segmentation model to learn the small region that matters is mostly a problem of imbalance.

Two pipelines

Both models are U-Nets: an encoder that shrinks the image while widening the features, a decoder that restores the resolution, and skip connections that carry detail across.

Liver. The liver model works on whole volumes. It takes the 2D U-Net and makes it three-dimensional: 3D convolutions and pooling, trilinear upsampling, and three output channels for background, liver and tumor. The original Decathlon set is over 26 GB, so the scans were resampled to 256 × 256 × Z, which brings it to about 2.5 GB. 105 subjects were used for training and 13 for testing.

Lung. Lung tumors are often very small, so the lung model is trained slice by slice with oversampling. A WeightedRandomSampler draws slices that contain a tumor more often, so that the network cannot settle on the trivial solution of outputting zero for every voxel. The lung model reaches a Dice score of 0.80 on validation data.

Both were trained on a single consumer GPU, a GTX 1050 Ti. The notebooks are public; the lung repository also includes rendered result pages.

What it led to

These projects are the practical foundation for a manuscript I am writing with Dr. Amineh Amini, A Systematic Mapping Study on Deep Learning-Based Biomedical Image Segmentation Techniques. It surveys architectures across biomedical imaging modalities and supervised, semi-supervised and unsupervised methods, with attention to evaluation metrics and clinical relevance, and sets out the open problems in 3D architectures and standardized evaluation protocols. It is in preparation.