The Molecular Imaging Technology Research Program (MITRP) is pleased to announce a new publication from one of our undergraduate researchers, Zhijin (Tracy) He. The paper was co-authored with MITRP Principal Investigator Dr. Alan B. McMillan. This research directly supports our group’s mission to advance imaging techniques and technology to gain new structural and functional information for the detection and assessment of human disease.
The study focuses on chest radiographs, more commonly known as chest X-rays. As one of the most frequently performed medical imaging examinations, chest radiographs are a cornerstone of diagnostic medicine. They provide a rapid, non-invasive view of the heart, lungs, airways, and bones, making them essential for diagnosing a wide range of conditions. Given their widespread use, developing automated tools to assist in their interpretation has the potential to significantly improve diagnostic accuracy and clinical efficiency.
Published in the Journal of Imaging Informatics in Medicine, the paper, “Comparative Evaluation of Radiomics and Deep Learning Models for Disease Detection in Chest Radiography,” addresses this opportunity by comparing two prominent AI methodologies: radiomics and deep learning.
- Radiomics is a method where computational models are guided to extract a large number of pre-defined, quantitative features from medical images, such as texture, shape, and intensity.
- Deep learning, in contrast, enables a model to learn directly from the raw image data, autonomously identifying the hierarchical patterns that are most predictive of disease.
A central focus of the paper is the critical importance of dataset size in the performance of these models. The study’s findings provide crucial insights into how each AI approach behaves under different data constraints.
Deep learning models are exceptionally powerful but also data-hungry. Their complex architectures contain millions of parameters that must be fine-tuned, and they require vast amounts of diverse data to learn the subtle patterns of disease without overfitting—a phenomenon where the model memorizes the training examples instead of learning generalizable rules. Tracy’s research confirmed that with larger datasets, the deep learning models scaled impressively, achieving the highest levels of diagnostic accuracy.
Conversely, because radiomics models are supplied with “handcrafted” features, their learning task is more constrained. This makes them less dependent on massive datasets to achieve strong performance. The results demonstrated that this approach is a reliable and effective alternative in settings where collecting thousands of patient scans is not feasible—a common challenge in medical research. The paper’s key contribution is providing this evidence-based guidance: the optimal choice between AI models is not just about which is theoretically superior, but which is best suited to the data resources available.
Tracy, a senior majoring in Statistics and Data Science, has been a valued member of our research team since joining in Fall 2022 through the Undergraduate Research Scholars (URS) program. Her hard work on this project exemplifies the significant contributions that undergraduate researchers can make to academic science!
Congratulations to both Tracy and Dr. McMillan on this achievement!
The full paper is available here: https://doi.org/10.1007/s10278-025-01670-9
Learn more about the Undergraduate Research Scholars (URS) program: https://urs.ls.wisc.edu