Biomedical image analysis: Segmentation, datasets, metrics and loss functions

Biomedical image analysis

The success of deep learning in image analysis has encouraged the biomedical imaging researchers to investigate its potential in analyzing various medical modalities to aid clinicians in faster diagnosis and treatment of diseases or infections like the on-going pandemic of SARS-CoV-2 (COVID-19). Following the deep learning usecases, the implication of classification can ascertain the presence or absence of disease in some modality e.g. the ground glass opacification (GGO) in the lungs via CT imaging. Furthermore, in localization, normal anatomy can be identified e.g. lungs in the CT or X-ray imaging, and later segmentation can generate refined boundaries around the GGOs to understand its impact on the anatomical structures for further analysis. Since, segmentation is an extension to classification, localization or detection, it offers very rich information about the disease and infected regions. With this interest, many architectures have been proposed for the segmentation of the targeted regions from vivid modalities. In addition, segmentation is the most widely researched application of deep learning in biomedical image analysis, where U-Net based segmentation architectures have gained significant popularity to develop computer-aided diagnosis (CAD) systems.

More details can be found in the following paper:

Punn, Narinder Singh and Sonali Agarwal. Modality specific U-Net variants for biomedical image segmentation: a survey. Artificial Intelligence Reviews (2022).

Summary of popular datasets for biomedical image segmentation

ISBI 2012Electron microscopy cell slides for cell segmentation
ISBI2D and 3D videos of moving cells for cell tracking
KDSB 2018Histopathological cell images for nuclei
PanNukeHistopathological slides for nuclei segmentation
DRIVERetinal fundus images for vessel
STARERetinal fundus imaging for blood vessel
CHASE_DB1Retinal fundus imaging for blood vessel
LiTSLiver CT scans for tumor segmentation
LIDC-IDRILung CT scans for cancer segmentation
LUNA 2016CT scans for lung nodule segmentation
xVertSegCT spine images for vertebra segmentation
SIIM-ACRChest X-rays for pneumothorax segmentation
ISICDermoscopy images for skin lesion segmentation
BraTS 2012 – 2020MRI modalities (T1, T2, FLAIR) for brain tumor segmentation
ISLESMRI scans for stroke lesion segmentation
ICCVBProstate MRI and retinal fundus imaging
IBSRRepository of MRI imaging
ACDC 2017MRI imaging for cardiac diagnosis and segmentation
PROMIS 2012Prostate MRI image segmentation
Medical Segmentation DecathlonMRI and CT modalities for tumor segmentation in various organs like liver, brain, lung, etc.
OASISMRI and PET images for aging analysis and segmentation
Head-Neck-PET-CTPET and CT imaging for tumor segmentation
BUSISUltrasound imaging for breast tumor segmentation
BUSIBreast ultrasound scans for tumor segmentation

Summary of performance metrics for biomedical image segmentation

Summary of performance metrics for BIS in terms of number of true positive (TP), true negative (TN), false positive (FP) and false negative (FN), predicted mask (\mathcal{P}) and ground truth (\mathcal{G}), \mathcal{H}(X,Y) is the directed AHD from X to Y with d as euclidean distance, \mathcal{V}_p and \mathcal{V}_g refer to the volumes of generated and reference segmentation.

Summary of performance metrics for biomedical image segmentation

In BIS, mostly the datasets are imbalanced i.e. the number of pixels/voxels concerning the target region (region of interest) are relatively less than the dark pixels/voxels (background region), due to which the metrics such as accuracy, which are best suited for a balanced distribution of data samples, are not recommended for BIS evaluation of the models. Among the shown metrics intersection-over-union (IoU or Jaccard index) and dice similarity coefficient are the most widely used evaluation metrics in BIS for various modalities. More details can be found here.

Summary of popular loss functions for biomedical image segmentation

Below table shows the summary of loss functions for biomedical image segmentation with respect to the predicted mask (\mathcal{P}) and ground truth mask (\mathcal{G}), \alpha and \gamma as constants, h is Hausdorff distance and d is the operator for Euclidean distance. More details can be found here.

Loss functions
Summary of loss functions for biomedical image segmentation

Leave a Reply