A new study has evaluated the performance of two state-of-the-art artificial intelligence (AI) models for cancer detection on mammograms, comparing their operation in countries with very different healthcare resources: Australia and Vietnam. The findings suggest that transfer learning can improve AI accuracy, although performance still depends on factors such as breast density and the technology used.
What was known
Population-based mammography screening programs in countries like Australia have been shown to reduce breast cancer mortality. However, such programs are not widely implemented in countries like Vietnam, where women also have high breast density and there is lower availability of expert radiologists. These differences raise questions about whether AI models developed in one context can perform equally well in another.
What this work adds
The study analyzed two datasets with the same number of mammographic cases (865 malignant and 865 normal) from Vietnam (Dataset 1) and Australia (Dataset 2). The researchers applied image enhancement techniques and used two AI models: the Globally-aware Multiple Instance Classifier (GMIC) and the Global-Local Activation Maps (GLAM). Breast density was classified by three radiologists into four levels and grouped into two categories (0%-50% and 50%-100%). The aim was to determine whether transfer learning and breast density affected AI performance.
Main results
The results showed that the GMIC model outperformed GLAM in terms of specificity and sensitivity in both datasets. Transfer learning significantly improved the performance of both models (p<0.05). In Dataset 2 (Australia), GMIC with transfer learning achieved 86.7% specificity and 88.2% sensitivity, surpassing the 80.4% and 81.2% obtained in Dataset 1 (Vietnam). Furthermore, the performance of GMIC with transfer learning in Dataset 2 was significantly higher in the 0%-50% and 50%-100% breast density ranges compared to Dataset 1.
What it means
This work underscores that AI is not a universal tool that works the same in all environments. Transfer learning, which allows a model trained in one context to be adapted to another, is emerging as a key strategy to improve detection in lower-resource countries. However, the differences observed based on breast density and data origin indicate that local validation and adaptation to each population’s characteristics are indispensable steps before implementing these technologies on a broad scale.
Reading from the breast unit
For a breast unit, this study reminds us of the importance of evaluating the performance of AI tools with local data and in the real clinical setting. Variability in breast density and imaging equipment can influence results, so AI implementation should be accompanied by rigorous local validation and continuous monitoring of its accuracy.
Reference: Jiang Z, Trieu PD, Barron ML, Lewis SJ. Investigating Artificial Intelligence Performance on Mammographic Cases in High- and Low-Resourced Countries. Asia-Pacific Journal of Clinical Oncology, 2026. https://doi.org/10.1111/ajco.70203
