Trying to understand a complex pathology report can be a frustrating and overwhelming ordeal for a patient facing a cancer diagnosis. But Rushitha Mamidala, a first-year doctoral student in USF鈥檚 Bellini College of Artificial Intelligence, Cybersecurity and Computing, wants to change that experience.
She鈥檚 developing an AI tool to support patients and their doctors by making cancer diagnoses more accurate and understandable.
鈥淐ancer jargon can be really complex for someone who is not aware of the medical terminology,鈥� Mamidala said. 鈥淲ith this tool, patients would see a more user-friendly version of their diagnosis and treatment plan, and clinicians would get an added level of validation.鈥�
Mamidala鈥檚 interest in AI development grew out of a natural language processing course she initially took at USF. It was taught by Assistant Professor Ankur Mali, who is now one of her co-advisors on the project.
鈥淚 didn鈥檛 know the depth of it when I was getting into it,鈥� she said of AI learning.
While Mali鈥檚 class sparked her curiosity, it was the work of Karen Panetta, who is the dean of graduate engineering education and professor with the Department of Electrical and Computer Engineering at Tufts University, who inspired Mamidala to pursue a project that had impact.
Panetta visited USF as a Computer Science & Engineering Distinguished Speaker in 2024, just as Mamidala was starting on her master鈥檚 thesis journey. There, Panetta discussed the challenges of AI accuracy and reliability. During that discussion, Mamidala was inspired by a story of how facial recognition had been used to reunite a child with their family after many years.
鈥淚鈥檝e always wanted to do something more impactful,鈥� Mamidala said. She found that in the health care arena.
鈥淗ealth, to me, is an area where the impact is direct. It affects the general public in a very real way.鈥�
A smarter, more efficient way to update medical AI
Mamidala鈥檚 work introduces a local RetoMaton framework that augments large language models (LLMs) with targeted, task-specific retrieval capabilities, enhancing their factual accuracy and making their reasoning process more transparent. She presented this novel framework specifically designed for mathematical reasoning at the 19th International Conference on NeuroSymbolic Learning and Reasoning.
鈥淟arge language models are powerful, but their decision-making is often a 鈥榖lack box,鈥欌�� Mamidala explained. They make predictions without clearly showing how they arrived at those results, clouding the transparency of the data.
By adding a structured form of external memory, which is tailored to a specific task 鈥� such as the uterine cancer data Mamidala is using 鈥� the RetoMaton framework enables explicit tracing of the information used in prediction, which improves factual reliability and model transparency. It also avoids the need to retrain the entire model every time new data is added, which can be expensive and time-consuming compared to traditional AI models.
LLMs and similar systems are typically trained on massive datasets. When researchers want them to learn new information, the models usually have to be retrained or fine-tuned. With an external memory system built into the automaton, the model doesn鈥檛 need to re-learn. A task-specific datastore can simply be plugged into the existing model. This adaptation doesn鈥檛 rely on new prompts or fine-tuning; instead, the RetoMaton operates directly on the model鈥檚 internal embedding space, dynamically tracking the structure of each sentence as it unfolds.
鈥淚t saves a lot of resources, like computing, power and energy,鈥� she said. 鈥淎ll we have to do is swap out the datastore. The model stays the same, but its answers adjust to the new task.鈥�
Partnering with clinicians to test real-world data
Mamidala is working with Dr. Adrian Kohut at Tampa General Hospital鈥檚 Cancer Center of South Florida, along with Mali and co-advisor John Templeton, an assistant professor at USF who is an expert in human-centered computing, particularly in the medical field. The advisors review and test the publicly available data before moving on to real patient data.
Mamidala鈥檚 work with Templeton and Mali is part of the college鈥檚 Smarter Health by Integrating, Enabling and Linking Devices (SHIELD) Lab and TKAI Lab. The SHIELD Lab focuses on developing smarter health by integrating, enabling and linking devices with projects in cancer, biomechanics, cardiology, neurology and more. The Trustworthy Knowledge Driven Artificial Intelligence Lab 鈥� TKAI Lab for short -- builds intelligent systems grounded in physics, information theory and human understanding.
鈥淩ushitha's work is significantly improving the way multidisciplinary teams collaborate in real time,鈥� Templeton said. 鈥淏y bringing expertise in NLP with an impressive ability to quickly understand complex health-related knowledge, she is mitigating communication gaps across clinical, research and technical domains. As a result, she is driving meaningful translational outcomes in a high-impact area 鈥� directly influencing how patients receive care, how physicians make decisions and how caregivers support ongoing treatment.鈥�
鈥淲e meet every week, and I鈥檓 learning a lot about cancer diagnoses,鈥� she said. 鈥淭here are so many details in a pathology report that can affect treatment. If we miss one, like metastasis, we have to go back and fix it. This model helps us make sure nothing gets missed.鈥�
Scalable tools for multiple cancer types
Being able to swap out that datastore will allow her to develop other tools to use for other cancer types. Working out the details of the AI model is key to its success and accuracy.
鈥淚 want to build something that makes a difference to people who are already dealing with so much,鈥� she said. 鈥淚f we can make this step in the cancer journey more understandable and less overwhelming, that鈥檚 worth it.鈥�