AI and big data in cancer treatment

Cancer, affecting one in three individuals in their lifetime, often requires complex treatment processes like radiotherapy. Using radiation to destroy tumors demands meticulous planning and considerable time. However, recent advancements in deep learning can potentially transform this field radically, making treatment planning more efficient by automating time-consuming tasks. Developing AI models capable of handling millions of parameters requires extensive data sets. In medical imaging, this is challenging due to strict GDPR and ethical regulations, making acquisition of large-scale data difficult.

The ASSIST project aims to optimize and simplify workflows in image-guided therapy procedures, with a key focus on addressing data accessibility challenges. Federated learning, with Scaleout as the underlying technology provider, is a central solution in this effort.

The ASSIST project has successfully secured ethical clearance for several Swedish hospitals, including those in Lund, Umeå, and Linköping, to participate in research experiments. These experiments focus on federated learning techniques using brain imaging data from radiotherapy treatments. This ethical approval marks a crucial step in the project, enabling the use of real-world medical data while adhering to strict privacy and ethical standards. Initial federated learning trials have already been conducted in collaboration with Linköping and Lund hospitals, using Scaleout's Studio platform.

This project also aims to assist physicians by reducing their manual tasks during these procedures, thus allowing them to focus more on patient care. The goal is to streamline physicians’ work, optimize imaging systems, improve patient outcomes, reduce human error, and lower costs.

Collaborative AI and synthetic images

Federated Learning offers a unique solution where hospitals can contribute to AI model training without sharing patient data. ASSIST leverages this method to build more robust datasets, enhancing radiotherapy planning. This approach allows us to gather extensive and diverse datasets while maintaining the privacy of medical images, in compliance with GDPR and ethical standards. It represents a step forward in advancing deep learning applications in medical imaging, particularly in the sensitive area of cancer treatment.

Complementing this, the second track of ASSIST focuses on the generation of synthetic images. These images, that closely resemble real patient scans, offer a potent augmentation to the Federated Learning approach. By integrating synthetic images into the dataset, we can significantly increase the volume of data available for training our models. This inclusion is important in refining the accuracy and robustness of our AI systems. By using Generative Adversarial Networks (GANs) and innovative diffusion models, we can create high-fidelity synthetic representations of brain tumors.

The goal of the projects extends beyond theoretical research; we aim to apply these models in real-world scenarios. Collaborations with various cities in Sweden and international partners like MIT and researchers in China and the Netherlands are crucial. These partnerships will enrich our training sets, offering a more diverse and comprehensive dataset.

Scene for AI at LiU

Listen to Anders Eklund from Linköping University, an ASSIST consortium partner, in a recent presentation going through the project in more detail.