Reteena Alzheimer’s Diagnosis Solution

Alzheimer's Disease (AD) is the most common type of neurodegenerative disease and the fifth leading cause of death among individuals aged 65 and up. AD is characterized by a gradual onset, with manifestations including memory loss, cognitive decline, and a variety of psychological and neurological issues (Yao et al., 2023).  Once a patient is diagnosed with AD there are no treatments that can effectively cure the disease, emphasizing the importance of an early diagnosis during the prodromal stage, known as mild cognitive impairment (MCI) to ensure actions can be taken to slow its progression. Traditional diagnostic techniques rely on the presence of specific symptoms such as long term memory loss or regressing of problem solving skills. However, these symptoms often don't show up until late stage symptoms occur. Neuroimaging can be used to achieve an early diagnosis because of the specific details regarding brain atrophy which these scans provide. The issue is that current neuroimaging techniques have large scan times, are expensive, and often confined to hospitals. This has led to a large number of AD cases, approximately 90%, in low income countries to go undiagnosed until late stage symptoms occur (Acuff et al., 2020)

Magnetic Resonance Imaging (MRI) is a neuroimaging technique used to obtain extensive information about the anatomical structures of the brain. It provides significant information regarding AD’s effects on brain structures including gray and white matter, ventricles, the amygdala, and the hippocampus. Scan segmentations is a common technique utilized by radiologists to analyze the volumetric changes within the brain structure (Smits 2021). An early diagnosis usually requires the meticulous analysis of patient data by radiologists and clinical experts. While this technique is effective, it is time-consuming, tedious, and subject to external bias. The use of deep learning to improve upon the segmentation of MRI brain scans is faster, more effective, and more accurate.

One major issue with common imaging techniques such as MRI is that they have very long scan times and require the patient to have access to a hospital with expensive equipment. One emerging technique within current research to address these problems is the use of low-field or low-resolution MRI scanners, which are composed of 50-100mT as opposed to common scanners which are 1-3T. This allows these scanners to be significantly cheaper, smaller, and have much shorter scan times as compared to traditional MRI scanners (Perron et al., 2023). These scanners can be transported to patients' homes and allow for an initial detection of AD without requiring a patient to have access to a hospital with expensive equipment. However, these scanners provide a significantly less amount of information when compared to expensive hospital-grade MRI scanners. This can lead to less detailed images and potentially missing subtle abnormalities, which can lead to misdiagnoses. While these scanners increase accessibility, they risk compromising diagnostic accuracy about a patient’s condition. Therefore, there is a need to improve the structural information that these scans provide for AD diagnosis. 

The first step of Reteena’s diagnosis framework was to synthetically create Low Field (LF) MRI images to train our models on. 

First, a fourier transformation converts the image from the spatial domain  into the frequency domain where the image is represented as a spectrum of various frequencies, with higher frequencies corresponding to finer details. Then, truncation is used to selectively remove higher frequency components of the image, effectively blurring the image by only leaving behind lower frequency/detail components of the image. Finally, an inverse fourier transformation is used to convert the lower resolution image back to the spatial domain, resulting in a blurred image with lower detail. These new transformed images had a much lower peak-signal-to-noise ratio (PSNR) and contrast-to-noise ratio compared to HF MRI scans. 

Reteena’s diagnosis technology is powered by artificial intelligence which is used for its two step enhancement and diagnosis. The first step of Reteena’s image enhancement pipeline increases the resolution of the LF scans through image super resolution in order to improve the quality and detail of the LF MRI scans, making the LF scans comparable to their HF counterparts. A Super Resolution Convolutional Neural Network (SRCNN) is used to achieve image super resolution (Dong et al., 2014).This technology ensures that despite the lower information provided in LF MRI images, the output closely approximates that of higher field strength MRI images in terms of detail and clarity, making it viable for accurate diagnosis.

The second step of Reteena’s image enhancement pipeline segmenst out the three regions of interest (ROIs) linked to AD (Amygdala, Hippocampus, and Ventricles) from the super resolved scans. UNET++ is used to segment out the ROIs because of its ability to delineate different objects from an image. The UNET architecture specializes in the accurate segmentation of medical images into various labels (Ronneberger et al., 2015) . The architecture is characterized by its "U" shape, consisting of an encoding path and a symmetric decoding path. Connecting the encoder and decoder are skip connections, directly flowing feature maps from layers in the encoder to their corresponding layers in the decoder to ensure precise delineation of complex features onto the image to be displayed as image labels, effectively segmenting the ROIs. The UNET++ architecture operates similarly to a traditional UNET but contains more skip connections to more effectively display the learned features as segmentations onto the scans. This structure allowed the UNET++ to excel in differentiating between the various tissue types and segment out the ROIs onto the super resolved MRI scans. 

The volumetric information of the ROIs (Amygdala, Hippocampus, and Ventricles) is then passed into a soft voting machine learning framework for an automatic diagnosis of AD. Three machine learning models (Linear Regression, Support Vector Machine, and Multilayer Perceptron) are utilized, where the majority consensus between the three is the final diagnosis.

Reteena Diagnostic Solution aims to provide a balanced approach between affordability and accuracy for individuals at high risk of Alzheimer's disease. In the landscape of Alzheimer’s diagnostics, two dominant approaches emerge. First, there are affordability-centric solutions that prioritize cost-effectiveness, making them accessible to a broader population. However, these solutions often sacrifice accuracy due to limitations in technology or resources. Individuals with financial constraints often rely on these options, but accuracy remains a concern. On the other hand, there are accuracy-centric solutions that offer high precision in detecting Alzheimer’s disease. Unfortunately, these cutting-edge methods tend to be expensive and confined to hospitals due to their size and large scan times. The dilemma persists: accurate diagnosis versus financial feasibility. 

Reteena seeks to offer a midpoint that provides to those who cannot afford or access detailed, accurate early-stage Alzheimer's diagnosis due to their financial status, allowing for timely interventions and improved quality of life. Early diagnosis of Alzheimer's disease is essential for effective treatment and management. When Alzheimer's is identified in the preclinical or Mild Cognitive Impairment (MCI) phase, when minimal symptoms are present, patients can receive timely medical, social, emotional, and financial support. Reteena empowers individuals at high risk of Alzheimer's to take proactive steps in managing their health, regardless of their financial circumstances. This early intervention through Reteena's diagnostic solution can enable better health outcomes and potentially slow the progression of the disease. By striking a balance between cost and precision, Reteena aims to democratize access to early Alzheimer's detection, enabling more people to receive the care and support they need, ultimately improving outcomes and quality of life.

Reteena Diagnostic Solution primarily serves healthcare providers, clinicians, and community organizations involved in the diagnosis and management of Alzheimer's disease. The solution leverages low-field or low-resolution, cheaper Magnetic Resonance Imaging (MRI) technology or MRIs with relatively low image resolution, particularly in underserved communities.

Our team's motivation stems from a deeply personal place: the fear of losing cherished memories to Alzheimer's disease. Alzheimer's disease is a devastating condition that can devour individuals of their most precious recollections. We understand the profound impact this condition can have on individuals and their families, which is why we are driven to find a solution that can preserve these invaluable memories. Financial difficulties should not be the barrier that exacerbates the burden of losing their invaluable memories. We believe that everyone deserves access to quality healthcare, regardless of their financial situation, and we are committed to ensuring that our solution is both accessible and affordable. 

Reteena was founded with a passion-driven mission to enhance medical diagnostics and ensure equitable access to quality healthcare. With a genuine intention of making a positive impact on healthcare, Reteena has experienced remarkable growth in a short period. Despite the inherent challenges and imperfections that come with being a young and relatively inexperienced team, Reteena has thrived in its pursuit of continuous growth and learning. Reteena consists of students who are not only passionate about and knowledgeable in the fields of AI, medical imaging, and Alzheimer's research, but also deeply committed to making a positive impact on healthcare. This multidisciplinary approach, combining individual interests and expertise, has been instrumental in our ability to develop innovative solutions and overcome challenges. Team Reteena has a strong enthusiasm for developing innovative solutions and is committed to improving healthcare accessibility and affordability. Additionally, Reteena has established partnerships with organizations and research institutions to ensure the successful implementation and adoption of our solution. These partnerships have provided us with access to resources, expertise, and opportunities for collaboration that have helped us advance our work. By leveraging these partnerships, we are confident that we can successfully implement and adopt our solution in healthcare settings. 

Our team is well-positioned to deliver this solution because of our passion, expertise, and commitment to making a positive impact on healthcare. We are driven by a shared goal of preserving memories and ensuring equitable access to quality healthcare for all. With our diverse skill set and strategic partnerships, we are confident that we can successfully develop and implement our solution, ultimately improving the lives of those affected by Alzheimer's disease.

Reteena is currently in the prototype stage of development. The company has created a functioning image resolution enhancement based diagnosis model, but it has not yet gone through the thorough scientific validation process using actual data. The model has been trained using the OASIS-3 dataset, which is still considered an early dataset for implementation in actual uses. Reteena is aiming to launch officially in the fourth quarter of the year after the continuous validation process is completed. Furthermore, Reteena has been continuously working on various grant and funding opportunities to make this scientific validation and research possible.

Reteena aims to deepen its involvement with the global community of health professionals specializing in artificial intelligence and diagnostic solutions through its participation in Solve. The main primary objective is to advance beyond the prototype phase, achieving validation and scientific approval for its solutions through rigorous research and testing. With the collaboration of professionals worldwide, Reteena aims to expedite this process, ensuring its innovative solutions meet the highest standards of efficacy and reliability in a short timeframe. Not unlike the retina in our eyes, which enables the vision of the limited world, Reteena seeks to enable the vision of the expansive understanding of Alzheimer’s disease. 

We believe in MIT Solve’s guiding mission: the right to access the full range of quality health services they need, when and where they’re needed. However, there are currently many inequities regarding medically underserved communities that we hope to help. With our diagnostic solution Reteena, we hope to help make an impact in both our local communities, but also worldwide. With our working prototype, we believe that we can help to improve health outcomes for Alzheimer’s Disease, which affects people living in poverty at significantly higher levels compared to people of a higher socioeconomic status. Allowing those specific underserved groups to have access to a comprehensive AI diagnostic solution to aid health professionals, will make large strides into this ongoing issue. 

We believe that MIT Solve allows us to effectively collaborate to improve diagnostic care on a worldwide level. Their grants would not only allow us to more efficiently and more accurately train our current prototype model, but they would also go towards a further understanding of the potential incorporation of AI into diagnostic testing. Embracing the newly created subfield of AI in medicine is a matter of the betterment of our medical landscape. Now more than ever, AI technology holds the key to unlocking a level of efficiency and accuracy that can preserve our medical industry, revitalize medical professionals, and save patients. Through this collaboration with MIT Solve, we believe that we can offer physicians a newly efficient method of diagnosing patients, taking the implementation of AI in stride, and leading us towards unparalleled advancements in the medical field.

Genuine Heart-Driven Changes is the culture code of Reteena. 

Enacting meaningful impact, especially in the highly specialized field of medicine, is truly arduous and perhaps improbable for students lacking extensive expertise and experience in this domain. Claiming that we will become the AI-powered diagnostic innovator without recognizing this reality may come across as unrealistic and beyond the bounds of rational contemplation. As a team, we do not wish to dream based on unforeseeable impossibilities, but rather one that dreams based on realistic possibilities while striving to push beyond perceived constraints with our authentic aspiration to make even a trivial impact. Reteena's Diagnostic AI, based on image resolution enhancement of low-field or low-resolution or low-resolution MRI scans, is not the flawless prototype that is perfectly ready to innovate the field. We are still grappling with a lack of experience, scientific validation, and errors that bombard our team at every turn. However, we are determined solvers, driven to overcome these challenges. With our genuine passion and enthusiasm, we are rapidly addressing the flaws and problems that arise. While our solution may currently not have an instantaneous impact in terms of actual usability and technological perfection, our concept and intention to enhance image resolution of low-field or low-resolution MRI scans for better accessibility and affordability could serve as a milestone in the better direction of AI usage to benefit the community. We acknowledge that our current prototype is far from perfect, but we are steadfast in our commitment to refine and validate our approach through rigorous testing and iteration. By leveraging our collective expertise and determination, we believe we can overcome the limitations we face and make meaningful contributions to the field of medical diagnostics. Our vision is not one of flawless innovation, but rather of steady, incremental progress driven by a genuine desire to improve healthcare accessibility and affordability. We are confident that our efforts will ultimately yield tangible benefits for the communities we aim to serve. 

We anticipate collaborating with experts from Seoul National University and KAIST in the coming months to scientifically validate their model and identify areas for improvement beyond the open-source data currently used. Additionally, despite the current model being built on a substantial amount of open-source data, we intend to progress towards using actual datasets by affiliating with local hospitals that utilize low-field or low-resolution MRI scans.

Our project aims to decrease the number of undetected cases of Alzheimer's in underdeveloped communities.  

A few of our impact goals include:

• Increase Accessibility: Our first goal is to increase the accessibility of early-stage Alzheimer's diagnosis in underserved communities. We aim to achieve this by partnering with local healthcare providers and community organizations to offer our diagnostic solution at affordable rates or through innovative payment models. We will measure our progress by the number of partnerships established and the increase in the number of individuals receiving early-stage Alzheimer's diagnosis in these communities.

• Improve Accuracy: Another goal is to continuously improve the accuracy of our diagnostic solution. We plan to achieve this by collaborating with experts in the field of AI and medical imaging to refine our algorithms and enhance the quality of our image resolution enhancement. Progress will be measured by comparing the accuracy of our solution against existing diagnostic methods and benchmarks.

• Raise Awareness: We aim to raise awareness about the importance of early-stage Alzheimer's diagnosis and the availability of our solution. This will involve conducting educational campaigns, participating in healthcare events, and engaging with the media. Progress will be measured by the reach and engagement of our awareness campaigns.

• Expand Reach: As we refine our solution and gain scientific validation, our goal is to expand its reach to other regions and countries facing similar challenges. While we’re currently headquartered in the United States, we’re already reaching out to other countries’ hospitals to create a partnership. We will measure this by the number of new partnerships established in different regions and the adoption rate of our solution in these areas.

• Ensure Sustainability: Finally, we aim to ensure the sustainability of our solution by establishing a viable business model that allows us to continue providing affordable and accurate diagnostics in the long term. This will be measured through our continued monitoring of our financial performance and adjusting our pricing and operations as needed to remain sustainable.

To measure our progress towards these goals, we will regularly track key performance indicators (KPIs) such as the number of diagnoses conducted, the accuracy of our solution, the reach of our awareness campaigns, the number of partnerships established, and our financial performance. By monitoring these KPIs and adjusting our strategies accordingly, we believe we can effectively measure our impact and continuously improve our solution to better support underserved communities.

Reteena’s image enhancement models were trained on 80 epochs with a batch size of 8 scans per iteration with an Adams Optimizer to reduce their respective loss function.

Reteena’s Super Resolution model, SRCNN, operates through a three-layered architecture designed for enhancing the resolution of images. The first layer, known as the patch extraction and representation layer, processes the input low-resolution image by extracting overlapping patches. These patches are then analyzed to identify and represent inherent features such as color, shape, and patterns, effectively converting spatial details into high-dimensional feature vectors. The second layer performs non-linear mapping, transforming these extracted feature vectors into a refined feature space that embodies the characteristics of a higher resolution. In the last stage, the reconstruction layer synthesizes these enhanced features to reconstruct the final super-resolved image. This layer meticulously combines the processed features to produce a high-resolution output from the initially low-resolution input. By training SRCNN with pairs of low and high-resolution images, it learns to predict the missing high-frequency details in low-resolution images, significantly enhancing the quality and utility of images produced by LF MRI scans. The SRCNN had mean squared error (MSE) as its loss function.

Reteena’s segmentation model, UNET++, consists of an encoder and decoder pathway. The encoder pathway operates similar to a traditional convolutional network, consisting of multiple convolutional layers responsible for understanding complex features within the image. The encoding pathway consists of the repeated application of two 3x3 convolutional layers followed by a rectified linear unit activation and a 2 x 2 max pooling operation to downsample the image. This is done by systematically reducing its dimensions through max-pooling layers while  learning complex patterns within the image using convolutional layers. This allows the encoding block to understand features in the scan such as color, shape, texture, and more to learn how to distinguish between the regions of interest (ROIs) and the rest of the image. The decoder block progressively reconstructs the image from the learned features using upsampling convolutional layers to enhance spatial dimensions back to the original size while overlaying segmentations onto the image. The UNET++ had dice loss as its loss function.

Three machine learning models are utilized for Reteena’s automatic diagnosis. The first model used is a linear regression which plotted the data in a 3D feature space according to the volume of its three ROIs and created the optimal linear hyperplane to separate the data into the two classes (AD and non AD). The second model utilized is a support vector machine (SVM) , which operates similar to a linear regression but is not confined to a linear separation and can more accurately separate the data into the two classes by creating a non-linear hyperplane between the data points. The final model utilized is a multilayer perceptron (MLP), which is a type of artificial neural network capable of learning how to separate the data into classes through complex learned patterns and relationships within the data. The final diagnosis is the majority consensus between the three models.

Five people work part-time on Reteena:
Seungyong Yang, Aarav Minocha, Jainish Patel, Jessica Wu, Bhavya Mamnani 

Less than 1 year

In our team Reteena, we are committed to fostering a diverse, equitable, and inclusive work environment where all team members feel welcomed, respected, and empowered to contribute. We do not artificially reinforce diversity but rather embrace differences as a natural part of our team dynamic. Our approach to diversity and inclusion is rooted in mutual respect and a shared commitment to our collective goals. We recognize that each team member brings unique perspectives, experiences, and ideas to the table, and we actively seek to create an environment where these differences are utilized to drive innovation and success. Rather than focusing on surface-level diversity metrics, we prioritize building a cu/lture of belonging where everyone feels valued for who they are. We encourage open communication, provide opportunities for professional development, and ensure fair and equitable treatment in all aspects of our work. By fostering this inclusive mindset, we can harness the full potential of our diverse team and deliver exceptional results for our organization. Diversity is not a box to be checked, but a fundamental aspect of our team's identity and a key driver of our collective success.

Reteena operates on a business-to-business model, wherein we have established partnerships with various institutions, including universities, medical clinics, and hospitals in underserved communities by seeking grants. We believe that healthcare should be a right, not a privilege. Due to this emphasis, we have created a verified 503(c)(3) nonprofit under fiscal sponsorship by the Hack Club. The U.S. Health Resources and Services Administration (HRSA) defines medically underserved areas/populations as areas or populations having too few primary care providers, high infant mortality, high poverty, or a high elderly population. According to a 2008 report from the World Health Organization, 90% of the world does not have access to MRI. Our first goal for the nonprofit is to provide free diagnoses to these regions and areas that cannot afford MRI scans. We host the diagnoses on our website and verify that these scans are coming from underserved medical institutions to ensure our server costs do not exceed the limit. We aim to ensure that Reteena is accessible to anyone.A large portion of people in the top 10% of the income bracket cannot afford MRI diagnosis. We charge them a small fee of $20, compared to the average diagnosis cost of $300. However, this cost is tentative and is decided on a case-by-case basis.

For large-scale hospitals, we are developing a custom API for healthcare developers, allowing them to access our technologies and cover our costs. Our main costs for Reteena come from model training, server hosting, and creating datasets. For the Alzheimer’s Diagnostic Tool, we utilized a portion of the OASIS-3 dataset due to its complexity and large size. We had no means of utilizing the dataset to its full potential due to limited computational power. We also had to switch from 3D MRI scans to 2D scans due to computational limitations. Our current AI model has high accuracy and recall rates, but these metrics could be higher if we utilized 3D scans. We diagnose Alzheimer's on a website that is run on a server by a private company. If our tool is used on a wide scale, we would exceed the server limits set by the private company and will need to switch to the paid tier. By earning monetary means from large hospitals, we ensure that our tools are up to standard, provide secured diagnoses, are extremely accurate, and deliver highly trusted results. We also gain monetary means through donations. Any monetary means that we deem unnecessary are donated back to the betterment of healthcare. We aim to provide research grants and donations to other nonprofits that align with our values to foster the open access of healthcare.

To be financially stable Reteena aims to be cost-efficient. If the AI model is cost-efficient, all the other costs are also lowered such as server hosting, model training, and inferencing. This overall reduces the diagnosis costs for customers who have access to healthcare but are not able to afford it. Our main costs surround model training since it requires heavy computational power. For example, our Alzhermier’s model was trained on 2D MRI scans, rather than 3D. It was trained on the Google Colab, eliminating the need for a physical, powerful computer. We did face hurdles due to GPU limits however but, by switching to 2D scans, it lowered training and inference times of scans. Another major cost that arises from Reetena, is diagnosis on the website. As we were finalizing our AI model we realized that we did not need a physical, powerful computer for server-hosting as our model was lightweight. We decided to use a private, external server hosting service and their free tier aligns with our needs. However, if our diagnostic tool is used on a wide scale, it will exceed the limitations on the free tier. We aim to charge the medical institutions or developers that aim to use our tool on a wide scale, allowing the cost to be directly invested back into our mission.

We recently partnered up with Hack Club for a fiscal sponsorship which allowed us to begin our non-profit. This has allowed us to operate globally and will continue to help us expand across nations. Their platform also allows all of our finances to be transparent to either only our team or to the public as well, to keep open communication about how the money would be properly distributed. They help to fund a year of our website costs as well as establish official 503(c)(3) non-profit status which has allowed us to kickstart our project. The current partnership also provides a platform to invite members to continuously grow our team. Currently, we have already garnered a lot of interest among our peers who wish to contribute and support the project.

We also plan to apply for multiple AI grants, especially through accredited universities. This would allow us to establish connections with professors who have experience in this field, as well as receive funding for our project.

However, that is only the beginning of what we hope to do. We plan to create fundraisers in our local communities to try and fund the resources that would be required to continue to optimize our training model. These would be conducted through events, school fundraisers, and restaurant partnerships. We also plan to accept donations through our website and advertise on social media platforms. Microgrants would be established so that we can reach out to the technological and medical fields with a pitch of our plan. By describing specific steps that we plan to take, along with what exactly their money is going towards, we hope to gain proper funding and become financially sustainable.