The problem HemoVeda: AI-Powered Anemia Detection Web App solves

HemoVeda empowers users and healthcare providers with a convenient, non-invasive tool for early anemia detection, particularly beneficial in regions with limited access to healthcare facilities. Individuals can use HemoVeda to gain immediate insight into their anemia risk by simply uploading images of their palms and conjunctiva, along with basic blood data. This approach provides a safe and accessible alternative to traditional diagnostic methods, which often require lab access and may delay timely intervention.

For healthcare providers, HemoVeda serves as a valuable resource for preliminary screening, supporting rapid decision-making and allowing for early intervention. By combining visual analysis of anemia symptoms with blood report data, the app delivers a more comprehensive and accurate risk assessment, reducing reliance on costly and invasive tests.

HemoVeda also facilitates continuous monitoring for those at high risk or with a history of anemia, allowing users to track symptoms over time. This is particularly useful for women in pregnancy, individuals in remote areas, or those unable to frequently access medical facilities. The app’s instant assessment feature saves time, allowing users and providers to address potential anemia risks promptly.

In addition, HemoVeda’s AI-powered analysis leverages state-of-the-art models to detect subtle changes indicative of anemia, making it a powerful preventative tool. This approach supports better health outcomes by enabling timely diagnosis, reducing the long-term physical impacts of untreated anemia, and potentially preventing more severe health complications.

Challenges I ran into

One significant hurdle we faced in developing HemoVeda was training the CycleGAN model using a ResNet-like architecture for the generator and a PatchGAN for the discriminator. The computational demands were extremely high, especially given our limited GPU access. Working within Kaggle’s workspace, we encountered several issues: GPU memory frequently maxed out, leading to out-of-bounds errors, and the model trained very slowly, further extending our project timeline. To address these challenges, we optimized data loading and experimented with model parameters to balance efficiency and performance, though the high computational cost remained a limiting factor.

Additionally, we aimed to include a live detection feature, where users could capture a real-time image of the palm region, automatically detecting the region of interest (ROI) for anemia assessment. Using OpenCV, we designed a method to identify the palm ROI accurately, but encountered challenges with camera quality, particularly on laptop webcams, which impacted image clarity and, thus, detection accuracy. Due to these limitations, we opted to exclude the live capture feature from this version, focusing on the manual image upload functionality while exploring potential enhancements for future releases.