MammoScreen | Devpost

Inspiration

Our team member's personal experience with losing his grandmother to cancer inspired us to build MammoScreen, an AI-powered solution for early detection of breast cancer. Our aim is to help medical professionals detect early signs of cancer in their infancy, empowering them to make informed decisions more swiftly and effectively.

What it does

MammoScreen utilizes machine learning algorithms to analyze images and identify potential indicators of breast cancer. Its primary function is to assist healthcare providers by augmenting their abilities, ultimately working towards saving lives through earlier interventions and better patient outcomes.

How we built it

We started by preparing a dataset consisting of 277,524 patches extracted from 162 whole mount slide images of breast cancer specimens. Then, we implemented a deep learning model using TensorFlow and Keras frameworks. We employed the VGG16 pre-trained model and fine-tuned it to suit our problem domain. Additionally, we experimented with various cross-validation strategies, such as Leave-One-Out Cross-Validation and Bootstrap Cross-Validation, to enhance the model's generalizability and reduce overfitting.

Challenges we ran into

Some challenges we faced throughout the development process include selecting appropriate cross-validation techniques for our imbalanced dataset, managing the trade-off between model complexity and overfitting, and using to best practices regarding privacy and security in handling sensitive medical data.

Accomplishments that we're proud of

Through careful design and implementation choices, we managed to achieve high accuracy rates in identifying both benign and malignant tissue regions. Moreover, our tailored cross-validation approaches helped improve the model's robustness and prevented overfitting despite the relatively smaller sample size.

What we learned

Working on MammoScreen allowed us to explore real-world applications of artificial intelligence and machine learning in the field of medicine. Specifically, we gained insights into employing transfer learning, tweaking pre-trained models, and implementing advanced cross-validation techniques to address the inherent challenges posed by medical imaging tasks.

What's next for MammoScreen

For future improvements, we plan to further refine our model by incorporating additional pre-trained architectures, exploring state-of-the-art segmentation techniques, and integrating clinical metadata to strengthen the precision.. Eventually, we aspire to deploy MammoScreen as a user-friendly web application accessible to healthcare practitioners worldwide, thereby contributing to improved breast cancer diagnosis and treatment.

Github link: https://github.com/heymukund07/BreastCancer_CNN-Model-/