Alzheimer's Detection Agent MRI Hippocampal Atrophy Analysis

\documentclass[11pt,a4paper]{article} \usepackage[utf8]{inputenc} \usepackage[T1]{fontenc} \usepackage[english]{babel} \usepackage[margin=2.5cm]{geometry} \usepackage{microtype} \usepackage{parskip} \usepackage{amsmath,amssymb,mathtools} \usepackage{graphicx} \usepackage{booktabs} \usepackage{caption} \usepackage{float} \usepackage{listings} \lstset{basicstyle=\ttfamily\small,breaklines=true,frame=single,numbers=left,numberstyle=\tiny,showstringspaces=false} \usepackage[hidelinks]{hyperref} \usepackage{xcolor} \usepackage[normalem]{ulem} \usepackage{soul} \usepackage{pifont} \newcommand{\cmark}{\ding{51}} \newcommand{\xmark}{\ding{55}} \newcommand{\markdownHeading}[1]{\section*{#1}} \newcommand{\mdsubheading}[1]{\subsection*{#1}}

\begin{document}

\begin{center} {\LARGE \textbf{Alzheimer's Detection Agent}}\[6pt] {\small By Sergio Rodriguez Mendoza and Laura Melissa Vargas Molina} \end{center}

\markdownHeading{Inspiration}

Our inspiration was born from a profoundly human challenge and a complex scientific problem: Alzheimer's disease. Millions of families worldwide watch as the memories and identities of their loved ones fade away because of this neurodegenerative disorder. In science, early diagnosis is the key to intervening and improving quality of life, but it runs into a wall: the initial signs of Alzheimer's are very subtle and are often confused with normal aging.

We were motivated by the idea of breaking through that diagnostic ``gray area.'' We knew that hippocampal atrophy (a small brain region crucial for memory) is one of the earliest physical indicators of the disease. Magnetic Resonance Imaging (MRI) can capture this atrophy, but its interpretation can be subjective. Therefore, we set out to create a tool that could translate those grayscale images into objective, clear data, offering a quantitative second opinion to help doctors and give patients hope.

\markdownHeading{What it does}

The Alzheimer's Detection Agent is a fully automated web platform that functions as a virtual neuroimaging analyst. Here's what it does:

\begin{enumerate} \item \textbf{Receives the Evidence:} A user uploads a \texttt{.zip} file containing a brain MRI in NIfTI format (\texttt{.hdr}/\texttt{.img}). \item \textbf{Initiates the Analysis:} The application extracts a strategic slice of the brain (slice number 64, which consistently shows the hippocampus) and subjects it to a rigorous image processing pipeline. \item \textbf{Isolates the Key Region:} Through a series of advanced filters and algorithms, the system precisely isolates and delineates the borders of the hippocampus. \item \textbf{Measures and Compares:} It calculates the exact area of the segmented hippocampus in pixels. It then compares this value against a clinically-informed reference threshold to determine if the size is within the normal range or if it shows signs of atrophy. \item \textbf{Delivers a Complete Report:} The platform presents a clear and easy-to-understand report that includes: \begin{itemize} \item A suggested diagnosis (\texttt{"Normal"} or \texttt{"Possible Alzheimer's"}). \item The exact calculated area of the hippocampus. \item An image of the segmented hippocampus for visual verification. \item A step-by-step gallery showing how the original image was transformed to reach the final result, ensuring complete transparency in the process. \end{itemize} \end{enumerate}

\markdownHeading{How we built it}

We built this project as an end-to-end solution, combining a modern user interface with a scientific analysis engine on the backend.

\mdsubheading{Data} We started with a dataset of anonymized MRI images. A crucial step was the normalization of all images to the Talairach and Tournoux coordinate space. This is like giving every brain a common map, allowing us to compare the structures of different individuals fairly and accurately.

\mdsubheading{Frontend} We developed a clean and responsive web interface using HTML, CSS, and JavaScript. The design is intended to be intuitive, allowing anyone, even without technical expertise, to upload a file and start the analysis with a single click.

\mdsubheading{Backend} The heart of our project is a robust API built with FastAPI in Python. This API manages the business logic and orchestrates the image processing pipeline.

\mdsubheading{The Tools (Image Processing)} This is where the work happens. We used a set of scientific Python libraries:

\begin{itemize} \item SimpleITK to load and handle complex NIfTI files. \item NumPy for all mathematical operations and image pixel manipulation. \item scikit-image and SciPy to implement the key pipeline steps. \end{itemize}

\mdsubheading{Key Pipeline Steps} \begin{itemize} \item \textbf{Contrast Enhancement:} Histogram equalization to make subtle details visible. \item \textbf{Noise Reduction:} Convolution and Gaussian filters to clean the image. \item \textbf{Edge Enhancement:} A Laplacian filter to sharpen boundaries of anatomical structures. \item \textbf{Precision Segmentation:} Active Contours (Snakes) and Chan--Vese for final refinement. \end{itemize}

\markdownHeading{Challenges we ran into}

\mdsubheading{The Subtlety of the Images} Brain MRIs are not like high-contrast photographs. The hippocampus often blends in with surrounding structures, making its delineation incredibly difficult. Finding the right combination of filters to enhance its edges without introducing artifacts was a process of intense trial and error.

\mdsubheading{Data Variability} No two brains are alike, and no two MRI scanners produce identical images. We had to create a very robust preprocessing pipeline that could handle these variations and effectively standardize the images.

\mdsubheading{Tuning the Segmentation Algorithms} Algorithms like Active Contours and Chan--Vese are powerful but also very sensitive to their parameters. We spent countless hours fine-tuning variables like alpha, beta, mu, and lambda to ensure the contours correctly adapted to the hippocampus in different patients, without ``leaking'' into other areas or failing completely.

\mdsubheading{Frontend-to-Backend Integration} Connecting a simple web interface with such a computationally heavy process on the backend was a challenge. We implemented asynchronous communication so the user receives real-time feedback while the server performs the complex calculations.

\markdownHeading{What's next for Alzheimer's Detection Agent}

\begin{itemize} \item \textbf{Large-Scale Validation:} Test the pipeline with a much larger and more diverse dataset to ensure robustness and reliability across different populations and scanner types. \item \textbf{Integrating Artificial Intelligence:} Incorporate deep learning models such as U-Net to improve segmentation accuracy and reduce processing time. \item \textbf{3D Volumetric Analysis:} Extend the algorithm from a single 2D slice to full 3D analysis to calculate total hippocampal volume, a more powerful biomarker than area alone. \item \textbf{Clinical Validation:} Collaborate with neurologists and radiologists to compare the tool's results with expert diagnoses, a necessary step to move from lab to clinical practice. \end{itemize}

\markdownHeading{About the Data}

The image dataset used in this project consists of brain MRI scans from multiple subjects. The original Siemens IMA proprietary files were converted to NIfTI1 16-bit format using a custom conversion program. During conversion, all direct identifiers such as subject IDs and scan dates were removed to protect participant privacy.

All scans underwent inter-scan head motion correction and were spatially transformed into the Talairach and Tournoux (1988) atlas space using a rigid-body transformation. A combined atlas, generated from a representative sample of young and older adults, was used to reduce bias when normalizing brains with varying degrees of atrophy. To optimize image quality, a 12-parameter affine transformation was applied and images were registered and resampled to an isotropic 1 mm resolution within the atlas space. Subsequent preprocessing steps included skull stripping with an adjusted mask and correction for intensity non-uniformity using a quadratic bias field model derived from phantom data.

Subjects were selected based on their Clinical Dementia Rating (CDR) to ensure the cohort included appropriate clinical groups for analysis. At the time this project began in 2024, the dataset was publicly available and accessible for research use. However, the dataset is not publicly available at present and access is currently restricted.

\markdownHeading{References}

\begin{thebibliography}{9} \bibitem{mayo2024} Mayo Clinic. Alzheimer's disease - Symptoms and causes. Accessed July 10, 2024. \url{https://www.mayoclinic.org/es/diseases-conditions/alzheimers-disease/symptoms-causes/syc-20350447}

\bibitem{izquierdo2013} Izquierdo C. and Medina J. Deterioro cognitivo leve y enfermedad de Alzheimer: diagnóstico precoz mediante neuroimagen funcional y estructural. Radiología. 2013;55(3):198-206. doi:10.1016/j.rx.2012.12.007 \end{thebibliography}

\end{document}