Dedicated Breast PET-Based Deep Learning Radiomics for Prediction of Pathologic Complete Response to Neoadjuvant
Tianhao Zeng1, Yilin He2, Teng Zhang1
1Jiangsu Key Laboratory of Intelligent Medical Image Computing, School of Artificial Intelligence, Nanjing University of Information Science and Technology, 219 Ning Liu Road, Nanjing 210044, China.
Cancers
|May 27, 2026
Summary
Dedicated breast PET scans show promise for predicting chemotherapy response in HER2-positive breast cancer. Fusing radiomics and deep learning features improved prediction accuracy, offering a potential noninvasive decision-support tool.
Area of Science:
- Oncology
- Medical Imaging
- Radiochemistry
Background:
- Accurate prediction of pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) is crucial for HER2-positive (HER2+) breast cancer treatment optimization.
- Noninvasive methods are needed to assess treatment response early, guiding therapeutic decisions.
Purpose of the Study:
- To explore the potential of baseline dedicated breast PET (D-PET) for predicting pCR to NAC in HER2+ breast cancer.
- To investigate a fusion strategy combining radiomics and deep learning features for enhanced prediction accuracy.
Main Methods:
- A multi-representation framework was developed using radiomics from metabolic subregions and deep learning on 3D tumor volumes.
- Intratumoral heterogeneity (ITH) was quantified as a comparator.
- Feature-level and decision-level fusion strategies were employed to integrate model outputs.
Main Results:
- Deep learning (AUC=0.79) and radiomics (AUC=0.78) showed comparable performance in predicting pCR.
- Fusion strategies improved prediction, with decision-level fusion achieving an AUC of 0.84 on test set 2.
- ITH analysis showed limited predictive value (AUC=0.61).
Conclusions:
- Baseline D-PET demonstrates promising potential for noninvasively predicting NAC response in HER2+ breast cancer.
- Fusion of deep learning and radiomics significantly enhances prediction performance compared to single models.
- D-PET models integrating these techniques could serve as valuable decision-support tools for clinicians.
