Accurate breast cancer risk prediction achieved with image conversion, AI-like classification

Computer scientists in the U.S. and China have demonstrated a mammographic image conversion method that incorporates optical-density features and combines with a computer-aided classification scheme to boost the accuracy of risk predictions for breast cancer.

Shiju Yan, PhD, of the University of Shanghai and four colleagues at the University of Oklahoma describe their work in a study published online in Oct. 3 in Academic Radiology.

Yan and team worked with an image dataset drawn from 168 negative mammography screening cases.

Their image conversion method involved re-rendering the patients’ original grayscale value (GV)-based mammographic images as optical density (OD)-based images.

For risk classification, they developed a two-stage approach tapping three artificial neural networks.

The first stage incorporated two artificial neural networks that the researchers trained using features computed separately from the GV and OD images of 138 patient cases, the authors explain.

The second stage brought in the third artificial neural network to “fuse” prediction scores produced by the two neural networks in the first stage.

The team tested risk-prediction performance on the remaining 30 cases at their disposal.

They found that, with the two-stage classification scheme, the computed area under the receiver operating characteristic curve (AUC) was  0.816 ± 0.071.

This was significantly higher than the AUC values achieved using two ANNs trained using GV features (0.669 ± 0.099) and OD features (0.646 ± 0.099) (P < .05).

“This study demonstrated that applying an optical density image conversion method can acquire new complementary information to those acquired from the original images,” Yan et al. conclude. “As a result, fusion image features computed from these two types of images yielded significantly higher performance in near-term breast cancer risk prediction.”

Among several limitations the authors acknowledge are their retrospective, lab-based data analysis and the small dataset they used, which is unlikely to replicate the diversity of data generated in a clinical screening environment.

“In our future study, we will continue to cooperate with doctors and collect new cases to expand the size of our dataset,” they write.

“The initial testing results are promising,” Yan and colleagues add. “However, the robustness of this new approach and risk prediction model (or classifier) needs to be further evaluated before it can be clinically acceptable to help establish a new optimal and personalized breast cancer screening paradigm in the future.”