Title of the project : « Generalized Saliency Prediction for the Segmentation of Biomedical Images »
Beginning : Oct. 2021 – Ending : Sep. 2024
Thesis director(s) : Hichem MAAREF (PR IUT Evry, IBISC SIAM team, 40%)
Co-advisor : Dominique Fourer (MCF Univ. Evry, IBISC SIAM team, 60%)
Team / Laboratory : SIAM – IBISC – Evry University / Paris-Saclay
School : Collège doctoral 580 – STIC
Contact : dominiqueDOTfourerATuniv-evryDOTfr
Keywords : Magnetic Resonance Imaging (MRI), image segmentation, tumor detection, deep transfer learning, adversarial training.
Abstract: The recent interest for artificial intelligence (AI) can mostly be explained by the success of deep learning techniques especially in computer vision. State-of-the-art methods based on deep neural networks can obtain good results when applied to biomedical images for tasks such as pattern recognition, object detection and segmentation. Unfortunately, all these methods require a large amount of goodquality annotated training examples to provide efficient prediction models. The specificity of biomedical applications, is that obtaining annotated is often complicated due to ethical issues about medical private information. Moreover, obtaining reliable annotations from specialists is often long and very expensive. Hence, this thesis addresses the problem of developing generalized saliency prediction models for image segmentation to deal with the problem of insufficient training dataset. Our aim is to develop and assess in real-world application scenarios (e.g. tumor segmentation), several learning paradigms related to Human perception, transfer learning and data augmentation.
Deep learning is beginning to have an impact in biological research and biomedical applications due to its capability to deal with large dataset, to extract new features and to consider the existing knowledge . However its efficiency strongly depends on the amount of available training data that can be complicated in very specialized tasks such as thrombus identification from stroke Magnetic Resonance Imaging (MRI) [9, 5]. Such tools can significantly help clinicians and regulators to make the good decision that could be decisive for saving the life of a patient.
Considering the problem of image segmentation from biomedical imaging, most of the state-of-the-art methods based on machine learning use task-specific annotated datasets. It result is a huge number of methods based on similar Deep Neural Network (DNN) models but trained using task-specific datasets . Moreover, clinicians which are the only ones responsible for medical decision making, require assistance tools to highlight the region of interest in an image.
The difficulty to obtain a sufficient amount of reliable task-specific biomedical training data for a supervised automatic approach requires to investigate new strategies. In fact, the physiological information of each patient can be secret or unavailable and the reliable annotations provided by a specialist can be long and expensive to obtain.
One solution suggested by very recent studies  proposes to develop new generic saliency functions based on the general behavior of radiologist which is able hierarchically organize the information in an image to identify a large amount of a pathologies from different kind of biomedical images with different modalities. The idea of extracting areas of interest from biomedical imaging has already shown its usefulness  and could benefit of being actualized using a deep neural network approach. Indeed, merging features to model the Human attention system  continue to gain interest with recent works such as  and begins of being exploited by early new deep learning architectures through attention-based models .
Hence, the goal of this thesis is to develop new deep learning paradigms leading to generalized biomedical image saliency function.
This thesis focuses on the biomedical image segmentation problem for which a large number of state-of-theart supervised methods based on DNN exist and can obtain efficient results with a sufficiently large training dataset. The main purpose here consists in dealing with the configuration where the number of annotated examples is insufficient to successfully train a DNN model. With this in mind, the PhD candidate will develop, combine and compare several new approaches to address this complicated configuration. Among the proposed solutions:
- Developing new generalized saliency models based on Human perception .
- Proposing new methods to efficiently extract common features used by different segmentation tasks through transfer learning .
- Investigating efficient data augmentation techniques to improve the robustness of models and to increase the size and quality of training dataset by considering adversarial training  and adversarial poisoning .
For each proposed approach, an effort is expected for analyzing and understanding the meaning of the relevant features and representations designed for estimating the areas of interest in a MRI.
- PhD Thesis of Jonathan Kobold (supervised by H. Maaref and V. Vigneron) on thrombus detection and segmentation from MRI (PhD defended in Feb. 2020) .
- Master thesis of Ikram Brahim (supervised by Fourer in collaboration with H. Maaref and V. Vigneron) on tumor segmentation from MRI using deep learning (Feb. 2019 – Sept. 2019) .
- Master thesis of Andrea Roncoli on biomedical image segmentation (Feb. 2019 – Sept. 2019).
- good machine learning and signal processing knowledge
- mathematical understanding of the formal background
- excellent programming skills (Python, Matlab, C++)
- good motivation, high productivity and methodical works
- an interest for AI and biomedicine
 Ikram Brahim, Dominique Fourer, Vincent Vigneron, and Hichem Maaref. Deep learning methods for mri brain tumor segmentation: a comparative study. In IEEE IPTA 2019, Istanbul, Turkey, November 2019. HAL: https://hal.archives-ouvertes.fr/hal-02296594
 Adrien Chan-Hon-Tong. Adversarial poisoning and inverse poisoning against deep learning. 2019.
 Yifan Chen, Han Wang, Xiaolu Sun, Bin Fan, and Chu Tang. Deep attention aware feature learning for person re-identification. arXiv preprint arXiv:2003.00517, 2020.
 Laurent Itti and Christof Koch. Feature combination strategies for saliency-based visual attention systems. Journal of Electronic imaging, 10(1):161–170, 2001.
 Kobold, V. Vigneron, H. Maaref, D. Fourer, M. Aghasaryan, C. Alecu, N. Chausson, Y. L’Hermitte, D. Smadja, E. Lang, and A. Maria Tomé. Thrombus segmentation on swan with multi-directional u-nets. In Proc. IEEE IPTA 2019, Istanbul, Turkey, November 2019.
 Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sanchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60–88, 2017.
 Polina Mamoshina, Armando Vieira, Evgeny Putin, and Alex Zhavoronkov. Applications of deep learning in biomedicine. Molecular pharmaceutics, 13(5):1445–1454, 2016.
 Takeru Miyato, Shin-ichi Maeda, Shin Ishii, and Masanori Koyama. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 2018.
 Oliver K Mohrs, Bernd Nowak, Steffen E Petersen, Matthias Welsner, Christine Rubel, Annett Magedanz, Hans-Ulrich Kauczor, and Thomas Voigtlaender. Thrombus detection in the left atrial appendage using contrast-enhanced mri: a pilot study. American Journal of Roentgenology, 186(1):198– 205, 2006.
 A Montoya Obeso, J Benois-Pineau, MS Garcia Vazquez, and AA Ramirez Acosta. Saliency-based selection of visual content for deep convolutional neural networks. Multimedia Tools and Applications, pages 1–24, 2017.
 Abraham Montoya Obeso, Jenny Benois-Pineau, Kamel Guissous, Valerie Gouet-Brunet, Mireya S Garcia Vazquez, and Alejandro A Ramirez Acosta. Comparative study of visual saliency maps in the problem of classification of architectural images with deep CNNs. In 2018 Eighth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1–6. IEEE, 2018.
 Chuanqi Tan, Fuchun Sun, Tao Kong, Wenchang Zhang, Chao Yang, and Chunfang Liu. A survey on deep transfer learning. CoRR, abs/1808.01974, 2018.
 Qiaorong Zhang and Huimin Xiao. Extracting regions of interest in biomedical images. In 2008 International Seminar on Future BioMedical Information Engineering, pages 3–6. IEEE, 2008.