Background:
Human Epidermal Growth Factor Receptor 2 (HER2) is a clinically actionable target in multiple cancers. HER2 tumor expression status has been traditionally categorized as broad binary classifications (i.e., positive or negative)[1] determined through protein or RNA-based assays on biopsied tissue, such as immunohistochemistry (IHC) with confirmatory in situ hybridization[2]. Prior to trastuzumab deruxtecan (T-DXd), most HER2-targeted treatments only benefited HER2-positive cancer patients. However, the success of T-DXd on a subset of HER2-negative (IHC 1+ and 2+) metastatic cancer patients [3] drives the clinical need for a non-invasive approach to granular HER2 status classification and T-DXd treatment response prediction. While HER2-targeted PET probes provide complementary information to IHC, standard uptake metrics have limitations in capturing biological heterogeneity[4].

Hypothesis/Objective:
This study evaluates the potential of trastuzumab-based HER2 PET radiomics to stratify HER2 expression status and predict tumor response to the antibody-drug conjugate T-DXd.

Methods:
A preclinical cohort (n=43 mice; 3D 64 Cu-trastuzumab PET scans) comprising multiple cancer types (breast, gastric, colorectal, pancreatic, and bladder) was analyzed for HER2 classification (IHC 0, 1+, 2+, 3+) based on annotated labels (Figure 1A). A separate syngeneic cohort (n = 10 mice; baseline 3D T-DXd PET scans acquired 24 hours post-treatment initiation) of HER2-low colon tumors was used to predict treatment response at day 23 following T-DXd initiation (Figure 1B). Responders were defined as achieving complete tumor regression by day 23, whereas non-responders exhibited no tumor shrinkage. In both analyses, a total of 107 established radiomic features were extracted from segmented tumors in each PET scan. Preprocessing steps of radiomic features included multicollinearity filtering and z-score standardization. Minimum Redundancy Maximum Relevance was the implemented feature selection framework. Subsequent principal component analysis (PCA) was applied for the treatment response prediction experiment. The predictive value of radiomics was evaluated via a K-Nearest Neighbors (KNN) machine learning model, under a repeated stratified 3-fold cross-validation (50 repetitions) framework for HER2 classification and Logistic Regression (LR) under a repeated leave-one-out cross-validation (10 repetitions) for treatment response prediction. To prevent data leakage, preprocessing steps and feature selection were fitted exclusively on the training folds of each cross-validation round (Figure 1C). The average cross-validated performance of radiomics over all repetitions was reported, and compared with the performance of the maximum and mean uptake values.

Results:
In the HER2 IHC classification task, radiomics achieved a multiclass Area Under the Receiver Operating Characteristic (AUROC) curve of 0.864 (95% Confidence Interval (CI) [0.853 – 0.876]) with most misclassifications between adjacent HER2 IHC categories. Radiomics significantly outperformed maximum uptake (AUROC: 0.780 (95% CI [0.766–0.794]; p =5.33 x10 -15) and mean uptake (AUROC: 0.742 (95% CI[0.727–0.757]; p= 1.07×10 -14 ). In the treatment response prediction task, radiomics also outperformed standard uptake metrics with an F-1 score of 0.922 compared to 0.571 and 0.615 for maximum uptake and mean uptake respectively.

Conclusion:
Radiomic features from HER2-targeted PET probes provided superior stratification of IHC expression levels and treatment response prediction compared to standard uptake metrics. These preliminary findings provide support for retrospective and prospective clinical studies exploring radiomics as a non-invasive precision approach to improve patient selection for HER2-targeted therapies.

Image/Figure:
Click to view full size
Click to view full size

Image/Figure Caption:
Figure 1. Experimental schema for HER2 classification and T-DXd treatment response prediction. A) HER2 IHC classification cohort: Mice with breast, gastric, colorectal, pancreatic, or bladder tumors (n = 43) underwent ⁶⁴Cu-trastuzumab PET imaging 18–24 h post-injection of tracer (6.512–9.916 MBq). Ground truth HER2 status was established via representative immunohistochemistry and scoring by a certified pathologist. B) Treatment response cohort: A syngeneic HER2-low colon cancer model (n = 10) received weekly T-DXd (10 mg/kg total). Baseline ⁶⁴Cu-T-DXd PET (13.838–14.985 MBq) scans were acquired 24 h post-treatment initiation. Response (Responder vs. Non-responder) was defined by tumor volume regression at day 23. C) Radiomics and modeling: 107 features were extracted from segmented PET volumes. Preprocessing included multicollinearity filtering, mRMR selection, and standardization. Models were evaluated using repeated cross-validation.

Author
Ayanda Grece Clara Huate
Washington University in St. Louis