Student Projects

4. ScientificRAG: An AI Agent Framework for DW-MRI Literature Mining and Automated Computational Experimentation

Project Context Retrieval-Augmented Generation (RAG)[4] and intelligent agents[5] offer new approaches for researchers to interact with scientific literature. These AI systems can process collections of academic papers, extract relevant information, and provide contextual answers while integrating with specialized computational tools. This technology has the potential to streamline literature review and computational analysis workflows, making them more efficient for researchers.

Project Overview: Develop an intelligent scientific research assistant that combines RAG capabilities with agent-based tool integration for diffusion MRI analysis. The system will ingest and index scientific papers on diffusion-weighted MRI, enabling researchers to query the literature naturally while seamlessly accessing computational tools like CACTUS[1] for substrate generation and MC-DC[2] for Monte Carlo simulations. The agent will understand research contexts, provide literature-backed answers, and automatically invoke appropriate computational tools based on user queries.

Learning Outcomes

• RAG Systems: Master document ingestion, vector storage, and retrieval pipelines for scientific literature
• Agent Architecture: Build intelligent agents that can reason about research questions and select appropriate tools
• Scientific Computing Integration: Connect AI systems with specialized research tools (CACTUS, MC-DC) via Model Context Protocol (MCP)[3]
• DW-MRI Domain Knowledge: Develop understanding of diffusion-weighted MRI through AI-assisted literature exploration
• Production AI Systems: Create robust, scalable systems for research environments

This project combines 50% AI/RAG coding, 40% research domain exploration, and 10% computational tool integration, providing hands-on experience with cutting-edge AI research tools.

Requirements

• Strong Python programming
• Interest in natural language processing and information retrieval
• Basic understanding of scientific computing workflows
• Knowledge in DW-MRI/biomedical imaging (advantageous)
• Curiosity about AI-assisted research methodologies

Technical Outcomes

• Production-ready RAG system for scientific literature analysis
• Intelligent agent capable of literature-informed computational tool selection
• Integration with CACTUS and MC-DC via MCP for seamless research workflows
• Documentation and deployment guidelines for research environments
• Demonstrated expertise in modern AI research assistance technologies

Supervisors: Dr Juan Luis Villarreal ([email protected]), Dr Jonathan Rafael Patiño ([email protected]) and Prof. Jean-Philippe Thiran

[1] Villarreal-Haro et al. “CACTUS: a computational framework for generating realistic white matter microstructure substrates.” Front Neuroinform, 2023.

[2] Rafael-Patino et al. “Robust Monte-Carlo simulations in diffusion-MRI: Effect of the substrate complexity and parameter choice on the reproducibility of results.” Front Neuroinform, 2020. 

[3]Singh, Aditi, et al. “A survey of the model context protocol (mcp): Standardizing context to enhance large language models (llms).” (2025).

[4]Gupta, Shailja, Rajesh Ranjan, and Surya Narayan Singh. “A comprehensive survey of retrieval-augmented generation (rag): Evolution, current landscape and future directions.” arXiv preprint arXiv:2410.12837 (2024).

[5]Sapkota, Ranjan, Konstantinos I. Roumeliotis, and Manoj Karkee. “Ai agents vs. agentic ai: A conceptual taxonomy, applications and challenges.” arXiv preprint arXiv:2505.10468 (2025).

---

5. Exploring age-related changes in cerebral white-matter streamlines at the segment-level in youth using open-access neuroimaging data

Figure. Illustration of the methods of Wasserthal et al., 2018

Background & Research Question:

Diffusion-weighted imaging enables in vivo mapping of white matter streamlines in the human brain, providing insights into how structural connectivity develops during childhood and adolescence. Previous studies (e.g., Reynolds et al., 2019, NeuroImage) have shown significant changes in white matter microstructure across this developmental period. However, these changes are typically assessed as averages across entire streamlines, which may obscure important regional variations along the streamlines.

In this project, we aim to go beyond whole-streamlines averages and investigate how white matter properties evolve along the length of individual streamlines. This fine-grained approach may reveal distinct developmental patterns occurring at different streamline locations, which could be linked to underlying neurobiological processes and cognitive outcomes.
In the context of a master project, a second stage will involve the investigation of the segment cauterization and correlation to the previously identified patterns. The student will be responsible to design and develop tailored strategies to perform tract specific segmentation of the brain streamlines.

Dataset:

We will use data from the Philadelphia Neurodevelopmental Cohort (PNC) (Satterthwaite et al., 2016), a large, publicly available dataset that includes neuroimaging and behavioral data from 700 participants with typical brain development aged 8 to 21 years.

Key features of the dataset for this project:

• Cross-sectional diffusion-weighted MRI data
• Preprocessed single-shell diffusion images
• Tractography-based segmentation using TractSeg (Wasserthal et al., 2018, NeuroImage; see Figure above) identifying 71 major white matter tracts
• Quantitative diffusion metrics—fractional anisotropy (FA) and mean diffusivity (MD)—sampled across 100 equidistant segments per tract

Objectives of the Project:

• Analyse how white matter properties (FA, MD) change with age along different segments of white matter tracts
• Identify segment-specific developmental patterns that may not be visible when averaging across whole tracts
• Apply appropriate statistical or computational approaches, such as functional data analysis or machine learning, to model these relationships

Requirement:

• Basic knowledge of statistics
• Interest in brain development and neuroimaging
• Motivation apply data analysis techniques (e.g., Python, R, or similar tools)

Supervision: Dr Jonathan Rafael Patiño ([email protected]), Vanessa Siffredi (UNIL/CHUV) and Prof. Jean-Philippe Thiran

---

6. Understanding brain’s White Matter organization and plasticity in children with AgCC using MRI tractography

Figure. T1-weighted, sagittal slice, of typically developing children, children with complete agenesis of the corpus callosum (AgCC) and partial AgCC. The corpus callosum is indicated with a yellow arrow in the typical development case.

Background & Research Question:

The corpus callosum is the largest white matter structure in the human brain, containing over 190 million axons that connect the left and right hemispheres. A developmental absence of this structure—known as agenesis of the corpus callosum (AgCC)—is one the most common congenital brain malformation (see Figure). AgCC can result in either complete or partial absence of the corpus callosum and is associated with variable cognitive, behavioral, and neurological outcomes. Diffusion-weighted imaging enables the in vivo reconstruction of white matter pathways. This neuroimaging tool is particularly well suited for studying populations with abnormalities of white-matter development —such as AgCC—as it enables the exploration of potential structural neuroplastic responses in such atypically developing brains. Despite its clinical relevance, research in this area remains limited, largely due to the rarity of the condition. A pioneering study by Bénézit et al. (2015) explored white matter reorganization in a small sample (n = 6, https://doi.org/10.1016/j.cortex.2014.08.022).

This project aims to replicate and extend the findings of Bénézit et al. using a larger and richer dataset, with the goal of better understanding white matter organization and plasticity in children with AgCC.

Dataset:

We will use data from a unique dataset, the “Pediatric Agenesis of the Corpus Callosum Project”, collected at the Royal Children’s Hospital in Melbourne, Australia (Siffredi et al., 2021 – 10.1093/cercor/bhaa289), which contained diffusion-weighted imaging of children with AgCC (n = 20) and typically developing children (n = 29) aged 8 to 16 years.

Key features of the dataset for this project:

• Cross-sectional diffusion-weighted MRI data of a unique clinical population
• Preprocessed two-shell diffusion images
• Whole brain tractography available
• Extensive neurobehavioral and clinical assessments

Objectives of the Project:

• Replicate the key findings of Bénézit et al. (2015) which will serve as a foundational framework for the initial analysis of white-matter streamlines in this population.
• Extend the analysis, depending on student interest, in one or more of the following directions:
  • Apply and compare different diffusion approaches to better characterise white-matter streamlines in such atypical brains
  • Assess symmetry and asymmetry in white matter streamlines across hemispheres
  • Explore atypical streamlines (e.g., sigmoid bundle)
  • Explore associations with clinical outcomes

Requirement:

• Basic knowledge of image processing
• Interest in neuroimaging and clinical neuroscience

Supervision: Dr Jonathan Rafael Patiño ([email protected]), Vanessa Siffredi (UNIL/CHUV) and Prof. Jean-Philippe Thiran

---

8. Deep-learning based characterization of lymph nodes tissue composition from histology images

Background & Research Question :

Lymph nodes are essential organs of the body’s immune system. They are one of the first organs to be reached by metastasis in case of primary cancer in another organ and their involvement is used by clinicians for cancer staging. In the context of our research within LTS5, we aim at deciphering the microstructural information encoded in diffusion-weighted magnetic resonance imaging (DW-MRI) signal of lymph nodes, to find biomarkers differentiating a metastatic microenvironment from a healthy one. To validate our method, we generate tissue-mimicking numerical substrates and compute their numerical DW-MRI signal using MCDC [1] [see Project 3 for more details]. The first step for this is having an accurate characterization of both healthy and metastatic lymph nodes composition in order to generate numerical substrates which match as closely as possible real lymph nodes tissues. Segmenting histology images of lymph nodes cuts provides a way to compute statistics about cells populations, density and metastatic invasion. 

The goal of this project is to use pre-existing methods to provide an enriched characterization of lymph nodes composition. The findings of this project will directly translate into the creation of numerical substrates closely matching human lymph nodes tissue. The results will therefore be part of our ongoing database matching numerical substrates to DW-MRI signal for subsequent publication. 

Dataset: 25 images of healthy lymph nodes and 30 images of metastatic lymph nodes are available.  

Objectives of the Project:

• Individual cells detection (optionally classification) using pre-trained model (e.g. Stardist [2]), finetuning if needed. 

• Region detection (optionally classification): primary and secondary follicles, germinal center, medulla, stroma, metastatic invasion, using pre-trained network (e.g. [3]), finetuning if needed.  

• Analysis and statistics generation: different cells population, sizes, fraction, density, tumor fraction, type of tumor. 

• Assessing limitations of histology and providing solutions. 

• Depending on the student interest and time: generation of tissue-mimicking substrates and generation of diffusion signal using MCDC [1] [see Project 3 for more details]. 

Requirement:

• Interest on human lymph node biology and pathology 

• Good programming skills (Python) 

• Image processing experience and deep learning 

• Basic statistics 

Supervision: Salomé Baup ([email protected]), Dr Jonathan Rafael Patiño ([email protected]) and Prof. Jean-Philippe Thiran 

References: 

[1] Rafael-Patino, Jonathan, et al. “Robust monte-carlo simulations in diffusion-mri: Effect of the substrate complexity and parameter choice on the reproducibility of results.” Frontiers in neuroinformatics (2020). 

[2] Weigert, Martin, and Uwe Schmidt. “Nuclei instance segmentation and classification in histopathology images with stardist.” 2022 IEEE International Symposium on Biomedical Imaging Challenges (ISBIC) (2022). 

[3] Khan, Amjad, et al. “Computer-assisted diagnosis of lymph node metastases in colorectal cancers using transfer learning with an ensemble model.” Modern pathology (2023).

---

10. Development of an Automated Staining Robot for Tele-Cytology in Developing Countries

Availability: Open to 1 student or a group of 2 students

Context:

Over the past decade, non-communicable diseases have become increasingly prevalent in Sub-Saharan Africa (SSA), particularly among women. Cervical cancer (CC) stands out as a major contributor to morbidity and mortality in the region. Globally, CC is the fourth most common cancer in women. In Cameroon, it is the second most common, with 2,770 diagnosed cases and 1,787 deaths reported in 2020 alone. As such, CC is the leading cause of cancer-related deaths among women in SSA.Despite being both preventable and curable, the implementation of primary preventive measures such as human papillomavirus (HPV) vaccination and regular screening remains challenging in many low- and middle-income countries (LMICs) due to resource limitations, infrastructure issues, and cultural barriers.One promising approach involves using self-obtained vaginal samples to test for HPV, the primary cause of cervical cancer. This method improves screening coverage and increases referrals for further examination. To maximize diagnostic accuracy, HPV testing must be complemented with cytology — the visual inspection of stained tissue samples under a microscope to detect cancerous cells.To address the shortage of trained personnel in LMICs, EPFL’s LTS5 laboratory, in collaboration with Geneva University Hospital (HUG) and local hospitals in Cameroon, has developed an AI-based tele-cytology solution. This allows expert pathologists in Switzerland to remotely access and analyze digitized tissue sample images, with AI assistance to highlight the most relevant regions.

However, a major bottleneck remains: preparing the cytologic samples for imaging. Specifically, the staining process — which involves immersing tissue slices in a precise sequence of chemical baths for specific durations — is time-consuming, repetitive, and prone to human error.

Project Proposal:

This project aims to design and build an automated staining robot capable of performing the entire staining process autonomously. One proposed design involves a carousel layout (image below), where multiple chemical baths are arranged in a circle and a robotic arm sequentially immerses the tissue sample into each bath for a programmed amount of time.

Project Steps:

1. Define technical specifications and constraints
2. Design mechanical components and prototype using 3D printing
3. Develop the control system using a microcontroller
4. Create programming software to set and adjust bath timings

Requirements:

• Experience with 3D mechanical design and 3D printing
• Knowledge of microcontroller programming and motor control

Supervisors: Alexandre Abbey ([email protected]), Dr. Pierre Vassilakos (HUG) and Prof. Jean-Philippe Thiran

---

11. Multimodal deep learning for cost-effective and accurate liver cancer detection from clinical Magnetic Resonance Images

Supervisors: Jonas Richiardi, Naïk Vietti Violi (CHUV), Jean-Philippe Thiran (EPFL)

INTRODUCTION
Hepatocellular carcinoma (HCC) is the second leading cause of cancer-related death worldwide [Torre2012]. The stage of disease at time of diagnosis determines prognosis and treatment effectiveness. In at-risk populations, clinical guidelines recommend semi-annual screening with ultrasound (US). Unfortunately, US sensitivity is low, around 47% for early-stage HCC [Tzartzeva2018]. Magnetic Resonance Imaging (MRI) has higher sensitivity (80% or more [Colli2006]), but at 30 to 40 minutes, full examinations are too long for clinical routine. Abbreviated MRI (AMRI) is a novel concept that consists in using a limited number of MRI sequences, thus reducing acquisition time (<10mins) and costs compared to MRI while keeping high diagnostic performance for HCC detection [Vietti Violi2020].

AI could be beneficial in the context of AMRI be reducing radiologist interpretation time and improving HCC lesion detection rate, particularly for small lesion – potentially curable. We have already developed a network for automatic HCC detection based on some MRI sequences that performs well on multiple datasets (80%+ sensitivity), including with different populations and contrast agents.
Here, our goal is to improve our existing deep neural networks for HCC detection, by integrating other MRI sequences that were not included in the initial network .

This is a challenging translational machine-learning project with potentially large and rapid clinical impact. It requires prior experience with Python. Experience in deep learning and image processing is not required but would be an advantage.

APPROACH
We will approach the problem as a multimodal semantic segmentation task, where the goal is to segment liver lesions (3D volume) using T1-weighted Dynamic Contrast Enhanced MRI (4D volumetric time series), T2-weighted images (BLADE/HASTE), and Diffusion-weighted images. As a baseline model, we will use our current attention U-net [Oktay 2018] with Tversky loss, a well-performing and robust convolutional network, together with our custom pre-training strategy.

After data preparation using our pipeline (co-registration of new modalities, including masking the liver, intensity normalization, etc), the work will focus on finding the optimal combination of modalities, tuning the hyperparameters of the new network, and modifying the training and architecture to handle missing modalities (modality drop-out). We will also include an objective to optimize the cost/efficacy ratio of the imaging procedures, taking into account costs for contrast agent and scan time, and estimate the energy expenditure of inference for various versions.

DATA AND EVALUATION
We will use open data for pre-training, and in-house data for fine-tuning, validation, and testing (363 patients of which 130 positives for HCC, each with 1-5 lesions), with dense annotations for the lesions. In addition, we will use data from other clinical partners to assess generalization ability of the model. We will evaluate performance using lesion-level metrics (segmentation: Dice coefficient, absolute volume difference, detection: sensitivity, specificity, AUROC, MCC) and patient-level metrics (sensitivity, specificity, MCC, PPV, NPV) for all-stage and early-stage HCC detection. In addition, we will benchmark the new approach with our own baseline and the nn-Unet [Isensee2020] with residual encoder.

REFERENCES
Colli et al (2006) Accuracy of ultrasonography, spiral CT, magnetic resonance, and alpha-fetoprotein in diagnosing hepatocellular carcinoma: a systematic review, Am J Gastroenterol 101(3) (2006) 513-23.
Isensee et al (2020) Automated Design of Deep Learning Methods for Biomedical Image Segmentation, arXiv:1904.08128v2
Oktay et al (2018), Attention U-Net: Learning Where to Look for the Pancreas, 10.48550/arXiv.1804.03999
Vietti Violi et al (2020) Gadoxetate-enhanced abbreviated MRI is highly accurate for hepatocellular carcinoma screening, European Radiology
Torre et al (2015) Global cancer statistics, 2012. CA Cancer J Clin 65:87-108
Tzartzeva et al (2018) Surveillance Imaging and Alpha Fetoprotein for Early Detection of Hepatocellular Carcinoma in Patients With Cirrhosis: A Meta-analysis. Gastroenterology 154:1706-1718 e1701

---

12. A Real-Time and Automated Acoustic Monitoring System for the Asian Hornet, Vespa velutina

The Asian hornet (Vespa velutina) is an invasive species that poses a major threat to both local biodiversity and beekeeping. In affected apiaries, it can decimate up to 30% of colonies, severely impacting pollination and honey production. Given the limitations of current monitoring methods, this project proposes an automated, real-time and non-invasive system relying on acoustic detection of hornets.

Project 12.1 : Asian Hornet Biometry Using Deep Learning

The current Support Vector Machine model is unable to determine whether successive detections correspond to the same individual or multiple hornets. This project aims to develop a deep learning model capable of identifying individual Vespa velutina hornets based on their acoustic signature. Such individual-level recognition would allow more accurate estimation of attack frequency and could be used to infer the distance of the hornet nest from the beehive, based on round-trip timing analysis. The project will involve dataset preparation, exploration of new feature extraction methods through literature reading, and training/testing of suitable deep architectures.

Student profile :

• Basic familiarity with Python and PyTorch or willing to learn
• Interest in signal processing
• Interest in apiculture : hornet audio samples will ideally need to be taken at apiaries, as the project will take place during the height of hornet activity.

Expected outcome:

A trained model able to differentiate individuals based on short audio samples, and a preliminary evaluation of its application for attack rate and nest distance estimation.

Supervisor : Prof. Jean-Philippe Thiran

Project 12.2 : Autonomous Power Supply Design for Asian Hornet Monitoring

The current prototype for detecting Vespa velutina near beehives relies on a direct power socket connection, which is impractical for deployment in remote rural areas where no such infrastructure exists. The goal of this project is to design and implement a self-sufficient power supply system that can operate the device continuously for at least one week. The student will begin by exploring various feasible solutions—such as solar panel integration, high-capacity power banks, or hybrid systems—while evaluating criteria like autonomy, cost and ease of maintenance. The second phase will involve implementing the most suitable solution and testing it under realistic operating conditions.

Student profile

• Basic knowledge in electronics and power supply design
• Interest in embedded systems and energy management
• Critical thinking

Expected outcome:

A functional autonomous power supply solution for the hornet detection device, allowing it to operate for at least 1 week independently.

Supervisor : Prof. Jean-Philippe Thiran