![](https://crypto4nerd.com/wp-content/uploads/2023/10/1X5ktpPhC09H3FGj2vM0DZg.png)
So you have trained your first segmentation model and are wondering how you can measure its performance. If you have not worked on deep learning segmentation models, here’s a 7-part series that will walk you through each step of the training, validation and inference process. In this post, I will discuss some common statistical evaluation metrics that can be used to measure your model’s predictive performance.
Essentially, biomedical image segmentation is an initial part of the diagnostic process where the “region of interest” is divided into different areas, separating the diseased region from adjoining healthy regions so a physician can interpret the results and make an assessment as to the disease severity and consider appropriate treatment options¹. Biomedical images come in different modalities such as X-ray Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET), Single-Photon Emission-Computed Tomography (SPECT), Ultrasound (US), PhotoAcoustic Imaging (PAI), Optical Imaging (OI) and Infrared (IR)/Raman Imaging².
In our deep-learning segmentation series, we trained a U-Net based model to segment Glioblastoma tumors from 3D MRI scans. We used the trained model to run inference on a set of test images that were not used for training. The model produced its version of the tumor segmentation map. To assess how well the model performed, we need to compare its segmentation map to the ground truth version, which in this case was produced by expert radiologists at the medical institution.
Let’s take patient 351 from our Glioblastoma dataset. Using ITK-Snap, we can open the main image of the brain and overlay the segmentation map on top. The left image is the ground truth segmentation, and the right image is the one predicted by the model.
A few things to note here. It’s a 3D image, with X, Y and Z axes. The image is Axially oriented, so its like looking down on the head from above. Horizontal slices are numbered from the neck region towards the top of the head…