This slide talks about my finding and summarization of the paper titled High-level Prior-based Loss Functions for Medical Image Segmentation: A Survey. Link to paper: https://arxiv.org/pdf/2011.08018.pdf
Quantum Computing: Current Landscape and the Future Role of APIs
High level priors in loss functions for Medical Image Segmentation Using CNNs
1. High Level Prior in
Medical Image
Segmentation
Submitted By:
Surabhi Govil
arXiv:2011.08018
[cs.CV]
2. Glossary:
● Summary of the paper
● Image segmentation using CNN
● Loss Function in CNN based image segmentation
● Priors
● High Level Priors
● Improving accuracy using High Level Priors in Loss Functions
● Drawbacks with using High Level features
● Conclusion
● References
3. Summary:
● The paper focuses on incorporating prior knowledge at loss function level to avoid
aberrations in medical images segmentations
● Compares and contrasts different types of high level prior information.
● The way by which knowledge about prior can be incorporated into segmentation
● Researches the work of other authors on priors.
● Limitations of current researches on priors
4. Image Segmentation:
● Segmentation is used in various applications including but not limited to medical image
segmentation, augmented reality, video surveillance.
● Image segmentation can be classified into instance based and semantic segmentation.
● Segmentation varies from classification in sense that classification categorizes image into
labels whereas segmentation identifies regions in an image.
Source: https://neptune.ai/blog/image-segmentation-in-2020
5. Types of Loss for Image Segmentation:
In neural networks, we seek to minimize the error. The objective function(set of weights) is often
referred to as a cost function or a loss function and the value thus calculated is referred to as
“loss.”
● Distribution loss: They aim to discern how are the classes distributed in a neural network.
The class imbalance, weight distribution etc.
● Region-based loss: These aim to minimize the mismatch or maximize the inter region
overlap between ground truth and predicted segmentation.
● Boundary-based Loss: Aim to minimize the distance between ground truth and predicted
segmentation
● Compound Loss: Summing of different losses like Dice+CE, Dice+TopK, Dice+Focal
In regression loss function is used to find the best fitting line that minimizes loss. By minimizing
the loss models accuracy can be increased.
7. What are Priors:
● Priors can be thought of as shape, size, topological features, range of values, of an object
that can distinctly identify it.
● In medical domain priors can be thought of as edges, range of values, shape of an organ,
tissue, tumour etc.
● Vessels, membranes, glands and other curvilinear objects segmentation can benefit from
incorporating topological priors constraints.
● Priors can be categorized into classes namely low - level or high level.
● Low level priors are derived from low-level features, such as the color, intensity, and
orientation. Ex: Gradient or distance maps
● High level priors on the other hand concern objects/ region in an image and their salient
features
9. High Level Priors:
● Features extracted from the object such as shape, size, appearance, location, range of
values etc. which help in interpreting it.
● In medical images these features can be thought of as being the topological and
geometrical constraints, image color intensity, etc of the organs/tissues.
● High level priors are features extracted from the object that can help in characterizing and
interpreting it.
● Optimization using high level priors can be for discrete hard label targets, or continuous
domains where loss is derived from soft probability maps.
10. High Level Priors in Loss function:
They enable the layers in the network to learn globally relevant discriminative features
which aid in estimating highly refined probability maps with lower loss.
● Object Shape
● Region based
● Topological priors
● Size of the feature
High level priors optimize loss by having a data fitting term with a constraint. Optimization
can be applied to continuous or discrete domain.
11. Drawbacks:
● Nature of prior:
○ This high level prior depends on the anatomy of object depicting the high level features.
○ This often requires domain knowledge and can be used in a semi-supervised or weakly learning
scenario.
● Soft probability maps:
○ Thresholding function required to make it a binary map and predict features.
○ This process can make loss function non- differentiable.
○ For other features discrete optimization techniques need to employed.
○ More complex the object more difficult to apply this method.
● Continuous vs discrete optimization strategies:
○ Continuous strategy adds the constraint in the main loss function as penalty.
○ Continuous optimization approach becomes unsatisfactory in case of multiple or competing
constraints.
○ Discrete optimization insures global optimum.
● Relationship between organs and loss design:
○ As we observed above topological priors look at the outer shape of objects thus work best for
curvilinear organs like the vessels, membranes, etc.
○ Various other priors exclusion & inclusion, inter-region all can be used to solve specific segmentation.
12. Conclusion:
● We have discussed the various types of high level priors the authors of the paper have
talked about.
● Prior based loss can be used for a weakly or semi-supervised learning
● Prior based loss can help generate anatomically more plausible regions.
● We saw with prior based loss computational complexity can remain reasonable.
● As authors discussed loss based priors allow to incorporate loss terms which cannot be
directly optimized by using SGD.
13. References:
● Chu He, Zishan Shi, Peizhang Fang, Dehui Xiong, Bokun He, Mingsheng Liao, "Edge Prior Multilayer
Segmentation Network Based on Bayesian Framework", Journal of Sensors, vol. 2020, Article ID
6854260, 11 pages, 2020. https://doi.org/10.1155/2020/6854260
● https://arxiv.org/pdf/2011.08018.pdf
● https://arxiv.org/pdf/2006.14822.pdf
● https://arxiv.org/pdf/1910.01877.pdf
● https://arxiv.org/pdf/1910.01877v1.pdf
● https://arxiv.org/pdf/1607.01092v1.pdf
● Guotai Wang, Wenqi Li, Sebastien Ourselin, Tom Vercauteren. "Automatic Brain Tumor Segmentation
using Cascaded Anisotropic Convolutional Neural Networks." In Brainlesion: Glioma, Multiple Sclerosis,
Stroke and Traumatic Brain Injuries. Pages 179-190. Springer, 2018. https://arxiv.org/abs/1709.00382