Locally advanced breast cancer (LABC) is typically treated with neoadjuvant chemotherapy (NAC) - chemotherapy given before surgery. The goal is to shrink the tumor to allow less extensive surgery and to test how well the cancer responds to treatment.
A key outcome measure is pathological complete response (pCR), defined as the absence of all invasive tumor in the breast and lymph nodes after completing NAC. Patients who achieve pCR have significantly better long-term outcomes, including improved disease-free survival and overall survival, especially in aggressive subtypes like triple-negative and HER2-positive breast cancer.
However, not all patients respond equally to NAC. Identifying which patients will achieve pCR before or early during treatment would allow clinicians to continue effective regimens for likely responders and switch strategies for non-responders, avoiding unnecessary toxicity from ineffective chemotherapy.
Current methods to predict pCR rely on clinical and pathological features such as tumor stage, grade, and molecular subtype, but no biomarker aside from molecular subtype has been validated for routine use. This study explored whether deep learning applied to pre-treatment 18F-FDG PET/CT images could predict which patients would achieve pCR.
This retrospective study included 31 patients with locally advanced breast cancer (stage II-III) treated with NAC between 2013 and 2020. The mean patient age was 54.26 years. All patients underwent primary breast surgery after completing NAC, providing definitive pathological assessment of response.
The cohort included a mix of tumor stages: predominantly T3 (17 patients) and T2 (11 patients), with the majority having N2 lymph node involvement. The mean SUVmax (maximum standardized uptake value) of primary tumors on PET/CT was 12.84, reflecting generally high metabolic activity. Molecular subtypes included HR+/HER2- (17 patients), HR-/HER2+ (7 patients), HR+/HER2+ (4 patients), and triple-negative (3 patients).
pCR was achieved in 9 of 31 patients (29%), while the remaining 22 (71%) had residual disease. NAC regimens used taxane-based and anthracycline-based combinations, with trastuzumab added for HER2-positive patients.
Pre-treatment PET/CT imaging was performed following a standard protocol using a Siemens Biograph mCT scanner. Images were exported as JPEG files, and tumor regions were manually cropped by specialist physicians to focus the deep learning model on the relevant anatomy. A total of 355 PET/CT image slices were generated from the 31 patients for model training and testing.
Convolutional neural networks (CNNs) are the dominant architecture for medical image classification. They learn hierarchical image features automatically from training data, without manual feature engineering. A key challenge in deep CNNs is the vanishing gradient problem, where gradient signals weaken during training as the network becomes deeper, degrading performance.
This study used ResNet-152, a residual neural network with 152 layers developed by He et al. ResNet addresses the vanishing gradient problem through shortcut connections (also called skip connections), which allow the input of a layer to bypass several intermediate layers and be added directly to later outputs. This enables the network to learn residual functions and supports much deeper architectures without degradation.
Each input PET/CT fusion image was resized to 224x224 pixels. The model applies multiple convolutional layers with batch normalization and ReLU activation functions, progressively compressing image representations to extract high-level features relevant to predicting pCR. A bottleneck technique reduces the 224x224 input to 56x56 to improve training efficiency.
The dataset was split into 80% training and 20% testing, and five-fold cross-validation was used to improve model reliability. This technique systematically rotates which portion of the data serves as the test set, reducing the risk of overfitting to any particular data split - important given the small sample size.
Across five-fold cross-validation, the ResNet-152 model achieved an overall accuracy of 84.79%, an AUC (area under the ROC curve) of 0.90, sensitivity of 85.2%, specificity of 84.8%, and F-score of 85.0%. These results indicate the model can reliably distinguish patients who will achieve pCR from those who will not based on pre-treatment PET/CT images alone.
Performance varied slightly across the five cross-validation folds. The best fold (C4) achieved 92% sensitivity, specificity, and F-score with a 98% AUC, while the weakest fold (C5) still achieved 80% sensitivity and 79% specificity. This variation reflects the inherent challenge of small sample sizes but the consistent performance across folds supports the model's generalizability.
Notably, SUVmax values were recorded but were not incorporated into the deep learning model - the model relied solely on the visual pattern information within the PET/CT image pixels. This is significant because it demonstrates that the deep learning approach extracts predictive information beyond the simple metabolic quantification captured by SUVmax alone.
Compared to other published methods, this model performed competitively. Previous radiomic feature-based approaches achieved AUCs of 0.70-0.73, while a combined radiomic-clinical model reached AUC 0.73-0.844. The ResNet-152 model's AUC of 0.90 is among the highest reported for pre-treatment pCR prediction from PET/CT in breast cancer.
The results build on a growing body of evidence that deep learning applied to PET/CT images can extract clinically relevant predictive information from breast cancer. Prior studies used traditional radiomics (mathematical feature extraction) or simple metrics like SUVmax, which require domain expertise and manual feature engineering and may miss complex spatial patterns.
Choi et al. demonstrated that a CNN model applied to PET/CT and PET/MRI images increased AUC from 0.652 to 0.886 at baseline compared to SUVmax alone, and from 0.687 to 0.980 for interim imaging. The current study's performance at baseline (AUC 0.90) is consistent with these results, suggesting that pre-treatment images alone contain substantial predictive information.
The clinical impact of accurate pCR prediction is significant. Patients unlikely to respond could be switched to alternative regimens or enrolled in clinical trials for novel agents, potentially improving their outcomes. Conversely, confirmed likely-responders could potentially forgo escalation of therapy. Accurate prediction could also reduce the economic and human costs of ineffective treatment courses.
The study has important limitations. The retrospective design introduces inherent selection bias, and the cohort of 31 patients is small - acknowledged by the authors as the most significant limitation. Sample size calculations were not performed prospectively. The absence of external validation cohorts limits generalizability, and the model did not incorporate quantitative PET metrics (SUVmax, TLG, MTV) or clinical variables that might further improve performance.
This study demonstrates that a deep convolutional neural network (ResNet-152) applied to pre-treatment 18F-FDG PET/CT images can predict pathological complete response to neoadjuvant chemotherapy in locally advanced breast cancer with 84.79% accuracy and 0.90 AUC. The model performs this prediction non-invasively before treatment begins, using only imaging data.
The findings support the broader potential of AI-driven imaging analysis in oncology decision-making. By identifying treatment responders before or early during therapy, deep learning tools could enable personalized treatment adjustments that improve outcomes and avoid unnecessary toxicity in non-responding patients.
Future work should address the current study's key limitations with larger, prospective datasets that include more patients across diverse breast cancer subtypes. Incorporating additional PET metrics such as total lesion glycolysis (TLG), metabolic tumor volume (MTV), and clinical variables into the model architecture may further improve predictive accuracy.
Multi-center validation studies with standardized imaging protocols will be essential to confirm that these results generalize across different patient populations and scanner platforms before clinical deployment. Integration with multi-modal data including genomics, circulating tumor DNA, and clinical biomarkers represents an exciting future direction for more comprehensive pCR prediction.