Final Project for ELEC4840
1 Overall
For the final project, each group is required to select one of the problems in
Assignment 2 (Skin lesion classification or CT segmentation) and then further
improve the selected task performance via two methods. You need to complete
the following tasks:
1. Run a baseline model on the provided dataset (you may use the code from
Assignment 2).
2. Propose or implement two methods to improve performance.
3. Write a summary report.
4. Give a presentation in the last class (16 May).
The specific requirements for the project are detailed in the problem statement.
1.1 Submission requirement
1. A report (.pdf, no more than 4 pages; Reference page do not count in.; Use
LNCS Template Link.).
2. Presentation slides (.pdf).
3. Code with the readme file.
1.2 Report Structure
1. Introduction: Introduce your topic and its importance; Give a brief literature
review based on the method you choose for improvement.
2. Methodology: Describe the methods you used for performance improvement.
Use figures and pseudo codes if necessary.
3. Experiment Results: Use tables and figures to present your results. Use a
few paragraphs to discuss your results.
4. Conclusion: Give a brief conclusion on your report.
1.3 Deadline for presentation and submission
PPT submission deadline: May 16, 11:59 pm
Report submission deadline (You can use the extra time after presentation to
further refine your written report.): May 20, 11:59 pm
Presentation time: May 16, 14:00 - 17:00
21.4 Grades
Grading policy: Presentation (30%) + Report (70%)
Presentation (30%):
– 20% for attendance
– 5% for well-prepared PPT, including (1) background introduction, (2) method
developments, (3) results analysis, and (4) conclusion.
– 5% for well-scheduled presentation: Each group member should present half
of the work.
Report (70%):
30%+30% for reproducing satisfactory results of two algorithms. 10% for a
well-written report.
An additional 20 points will be given for any novel ideas and improvements
you have made to improve the performance. To obtain it, please clearly indicate
what you have tried and what results you have achieved in your report.
2 Topic Selection
You can select one of the problems in Assignment 2 (Skin lesion classification
or CT segmentation) and then further improve the task performance via two
methods. You can use the datasets provided in Assignment 2 and the codes you
have implemented to finish the project.
3 Possible Directions of Enhancement
Here are some improvement directions that can be applied to both classification
and segmentation tasks for your reference. Please notice that you might need to
modify some of these method’s codes to adapt to your own problem.
3.1 Semi-Supervised Learning
In addition to the fully labeled data provided in the previous assignment, you
can also leverage more unlabeled data to expand your data set for a further
performance boost. Here, we provide additional unlabeled data for skin lesions
(link) and CT images (link) for you to use. Please indicate what and how many
unlabeled data have you used in your final report.
You can implement one of the semi-supervised learning algorithms intro-
duced in our class to enhance the performance. The possible solutions are: (1)
Classification: Pseudo-Label [6], π model [5], temporal ensembling [5], mean
teacher [7], (2) Segmentation: UA-MT [11], UMCT [9], and CPS [3]. In addition
to these methods, you can also try to propose some new ideas to improve the
performance.
Final Project for ELEC4840 3
3.2 Domain Generalization
The model is likely to suffer performance degradation if the source of the data is
diverse. For example, the data is collected from different hospitals, and different
hospitals may collect the data using different devices on different patient cohort.
Therefore, the model cannot generalize well to these different novel cases.We can
leverage domain generalization to mitigate this issue.
You can try to implement one domain generalization (DG) methods intro-
duced in class (FACT [10] for classification task, Dofe [8] for segmentation task
or other methods) In addition to these methods, you can also try to propose
some new ideas to improve the performance.
3.3 Contrastive Learning
Contrastive learning is commonly used to teach models to encode representations
by maximizing the distance between differently augmented views of the same
data point while minimizing the similarity between representations of different
data points, aiming to leverage the inherent structure within the data itself. It
is also effective in scenarios where labeled data is scarce.
Several works have explored the essence of contrastive learning and proposed
efficient algorithm, such as MOCO [4] and SimCLR [2] for classification and [1]
for segmentation. You can refer to those researches and apply contrastive learn-
ing to improve your model.
References
1. Chaitanya, K., Erdil, E., Karani, N., Konukoglu, E.: Contrastive learning of global
and local features for medical image segmentation with limited annotations. Ad-
vances in neural information processing systems 33, 12546–12558 (2020)
2. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con-
trastive learning of visual representations. In: International conference on machine
learning. pp. 1597–1607. PMLR (2020)
3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation
with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition. pp. 2613–2622 (2021)
4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised
visual representation learning. In: Proceedings of the IEEE/CVF conference on
computer vision and pattern recognition. pp. 9729–9738 (2020)
5. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv
preprint arXiv:1610.02242 (2016)
6. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning
method for deep neural networks. In: Workshop on challenges in representation
learning, ICML. vol. 3, p. 896 (2013)
7. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged
consistency targets improve semi-supervised deep learning results. Advances in
neural information processing systems 30 (2017)
48. Wang, S., Yu, L., Li, K., Yang, X., Fu, C.W., Heng, P.A.: Dofe: Domain-oriented
feature embedding for generalizable fundus image segmentation on unseen datasets.
IEEE Transactions on Medical Imaging 39(12), 4237–4248 (2020)
9. Xia, Y., Yang, D., Yu, Z., Liu, F., Cai, J., Yu, L., Zhu, Z., Xu, D., Yuille, A., Roth,
H.: Uncertainty-aware multi-view co-training for semi-supervised medical image
segmentation and domain adaptation. Medical image analysis 65, 101766 (2020)
10. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for
domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition. pp. 14383–14392 (2021)
11. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling
model for semi-supervised 3d left atrium segmentation. In: Medical Image Comput-
ing and Computer Assisted Intervention–MICCAI 2019: 22nd International Confer-
ence, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22. pp. 605–613.
Springer (2019)