BainGlobe Project Proposal

Cell Classification in BrainGlobe Proposal-Google Summer Of Code 2025

				Personal Details

Name :Karnakota Nikhil Kumar Country of Residence : India Timezone : IST(India) Typical Working Hours : 11:30-13:30, 15:00-19:00, 22:00-05:00 ( IST ) 6:00-8:00, 9:30-13:30, 16:30-23:30 (UTC) Email : [email protected] Github : https://github.com/Nikhilkumarr-dev Zulip : https://neuroinformatics.zulipchat.com/#user/885338 Language : English

			Project Proposal

Project Title : Cell Classification in BrainGlobe

Overview : The BrainGlobe project is used for the classification and processing of brain images to identify cells. It facilitates Python-based tools for computational neuroanatomy and aims to develop interoperable tools in Python. The cellfinder tool within BrainGlobe is designed for automated 3D cell detection in large-scale whole-brain images. It uses traditional image processing to identify potential cell candidates and then classifies them using a ResNet-based deep learning model. However, deep learning architectures have significantly evolved since ResNet, and this project aims to explore newer, more efficient architectures to improve classification accuracy, speed, and computational efficiency. As part of this project, I will develop and implement at least one advanced deep learning model (such as EfficientNet, DenseNet, or Vision Transformers) in Python using PyTorch/Keras and integrate it into cellfinder. A detailed quantitative comparison will be conducted between the newly implemented architecture and the existing ResNet model to measure performance improvements. Additionally, I will ensure the implementation is well-tested, documented, and easy to use within cellfinder’s ecosystem, including its core module and Napari plugin. Beyond the technical implementation, I will contribute to comprehensive testing, documentation, and community engagement, ensuring that developers and researchers can seamlessly adopt the new architecture. To make the research accessible, I will write a detailed blog post explaining the architecture, its advantages, and challenges compared to ResNet.

Previous Projects Around Collaboration : Risk-Sensitivity in Multi-Armed Bandits : Surveyed and implemented risk-sensitivity methods for stochastic bandit problems, and upgraded the Explore-Then-Commit algorithm for VaR and cVaR measures with competent performance.

Leveraging Ontological Knowledge for Neural Language Models :Incorporated Weight Initialization in learning word embeddings using the WordNet Ontology for a task in the Construction domain, resulting in a faster convergence rate and better representation of domain-specific terms

Skin Disease Diagnostic System : Designed a web application that attempts to diagnose skin diseases based on images of the user’s skin powered by a deep neural model trained on a dataset created by scraping images from the web Skin Disease Diagnostic System Microsoft code.fun.do Contest, Indian Institute of Technology Madras Designed a web application that attempts to diagnose skin diseases based on images of the user’s skin powered by a deep neural model trained on a dataset created by scraping images from the web

				Schedule

Milestones

1. Literature Review and Dataset Exploration a. Study existing deep learning architectures used in cell classification. b. Explore BrainGlobe’s existing dataset and understand preprocessing requirements. c. Identify gaps in the current cellfinder model that can be improved.
2. Experimenting with New Deep Learning Architectures a. Selecting potential architectures. b. Train baseline models using existing cellfinder data. c. Compare performance with current cellfinder model
3. Model Optimization and Hyperparameter Tuning a. Apply techniques like dropout, batch normalization, and learning rate scheduling. b. Optimize model size and inference time for better performance. c. Conduct ablation studies to understand which components contribute most.
4. Integration with Cellfinder Pipeline a. Modify the existing cellfinder codebase to integrate the new model. b. Ensure compatibility with BrainGlobe’s ecosystem. c. Run end-to-end tests with real microscopy data.
5. Evaluation and Benchmarking a. Compare the new model’s accuracy, speed, and robustness with existing models. b. Perform cross-validation to verify generalization. c. Document results and prepare reports.
6. Documentation and Community Feedback a. Write detailed documentation on model improvements and integration. b. Get feedback from the NeuroInformatics Unit Community and make final adjustments. c. Submit the final proposal for integration into the main cellfinder repository.

Research Milestones in the Future Exploration of Transformation based model : Investing the potential of vision transformers and self attention mechanisms for improved cell classification. Supervised Learning : Developing methods to leverage microscopy data to improve data to improve model generalization Multi-Modal Data Integration : Enhance classification accuracy by integrating multiple imaging modalities

				Timeline

09th April - 08th May: Getting Familiar with the Codebase a. Explore BrainGlobe cellfinder repository. b. Understand existing deep learning models used in the project. c. Identify key areas for improvement. 8th May - 31st May: Community Bonding & Architecture Selection a. Interact with BrainGlobe mentors and developers. b. Finalize the deep learning architectures to test. c. Set up the training pipeline for experimentation. 1st June - 25th June: Initial Experiments a. Train baseline models with existing cellfinder data. b. Compare initial results with the current cellfinder model. c. Identify areas where the new models outperform or underperform. 26th June - 10th July: Optimization & Hyperparameter Tuning a. Tune parameters to improve accuracy and efficiency. b. Test model generalization with different datasets. c. Validate results with domain experts. 10th July - 18th July: Integration into Cellfinder a. Modify the cellfinder pipeline to include the new model. b. Ensure smooth inference and training workflows. c. Test the model in real-world conditions and Midterm evaluation 19th July - 31st July: Performance Benchmarking a. Compare inference speed and accuracy with existing models. b. Optimize model size for deployment efficiency. c. Generate benchmark reports for evaluation. 1st Aug - 15th Aug: Final Adjustments & UI Enhancements a. Improve visualization tools for better interpretability. b. Ensure smooth integration with BrainGlobe’s tools. c. Address any remaining issues or mentor feedback. 16th Aug - Final Evaluation: Testing & Documentation a. Conduct extensive testing with diverse datasets. b. Document all findings and create user guides. c. Submit the final implementation and research report.

Technical Requirements Operating System: Linux or macOS , Windows with WSL Python Version: Python 3.8+ Virtual Environment: Conda or venv for dependency management Deep Learning Frameworks: PyTorch, Keras Scientific Computing: NumPy, pandas, SciPy Image Processing: OpenCV, scikit-image Data Handling: h5py (for handling large datasets), PIL (Pillow for image loading) Visualization: Matplotlib, Seaborn, TensorBoard (for model performance tracking)

Post-GSoC Commitment I plan to continue contributing to the NeuroInformatics Unit even after GSoC ends. My goal is to remain engaged with the community by fixing bugs, improving documentation, and assisting new contributors. Additionally, I will explore more challenging features to enhance my learning and provide long-term value to the project. By staying active, I aim to deepen my expertise and strengthen my relationship with the open-source community at NeuroInformatics Unit.

Open Source Development Experience

I actively participated in Hacktoberfest, where I contributed to open-source projects by submitting pull requests, improving documentation, and engaging with the community. This experience introduced me to the fundamentals of open-source collaboration, including working with Git, GitHub, issue tracking, and pull request workflows. Additionally, I have been exploring open-source projects related to machine learning, deep learning, and web development.

Technologies known

Programming Languages: Python, JavaScript, TypeScript, Java, C Web Development: React, Next.js, Node.js, Express.js, HTML, CSS, EJS, jQuery State Management: Redux Toolkit, Recoil Databases: MongoDB, MySQL, PostgreSQL Machine Learning & Deep Learning: PyTorch, TensorFlow, Keras, NumPy, pandas, OpenCV, Scikit-Learn Cybersecurity & DevOps: Linux, Ethical Hacking, Threat Intelligence, Malware Analysis Version Control & Open Source: Git, GitHub, GitLab Other Tools & Libraries: Jest (Testing), bcrypt, Zod, WebSockets, Figma, Canva

Personal Statement

Past Experience: 1.Winner of two hackathons, where I developed innovative solutions using web technologies and machine learning. 2.GDSC Hackathon: Built a health-tracking application using React, Firebase, and MongoDB. 3.Department Hackathon: Developed a school loan system using Next.js and MongoDB. 4.Participated in Hacktoberfest, contributing to open-source projects. 5.Worked on Rocket.Chat and Postman as part of my open-source journey. 6.Contributed to various GitHub repositories, focusing on MERN stack and cybersecurity-related projects.

Motivation Why this project : I am motivated to work on this project because it aligns with my passion for deep learning, image processing, and contributing to open-source advancements in computational neuroanatomy. Cellfinder is a crucial tool for automating 3D cell detection in large brain images, and improving its deep learning architecture can significantly enhance its accuracy and efficiency. My background in machine learning, deep learning, and frameworks like PyTorch, TensorFlow, and Keras makes this an exciting challenge for me. The opportunity to explore newer neural network architectures beyond ResNet and implement them in a real-world biomedical application is both intellectually stimulating and impactful. Additionally, working on this project will allow me to contribute to an established open-source ecosystem, collaborate with experienced mentors, and gain hands-on experience in optimizing deep learning models for large-scale image classification.

Match Why you : I am a passionate deep learning and machine learning enthusiast with experience in PyTorch, TensorFlow, and Keras. My background in image processing, along with my open-source contributions and participation in hackathons, makes me a strong candidate for this project. I have a keen interest in optimizing neural networks and improving real-world applications. This project aligns with my skills and aspirations, allowing me to contribute meaningfully while learning from experienced mentors.

Availability : I have no commitments in the summer. I’ll be staying back home for the most part of it. I have mentioned my typical working hours above and on an average will be able to spend 40 hours per week on the project.