Breast cancer is not a single disease but a collection of biologically distinct subtypes with different prognoses and treatment responses. The two most common forms are invasive ductal carcinoma (IDC) - the most prevalent, originating in the milk ducts - and invasive lobular carcinoma (ILC) - the second most common, accounting for 5-15% of all breast cancers. ILC tends to be hormone receptor-positive (ER+/PR+) and HER2-negative, has ill-defined margins, and exhibits distinct metastatic patterns including spread to gastrointestinal and peritoneal sites.
Accurate subtype identification is critical because treatment regimens, prognosis, and monitoring strategies differ substantially between subtypes. Key prognostic factors include tumor grade, hormone receptor (ER/PR) and HER2 status, lymph node involvement, and extent of metastatic spread. Determining axillary lymph node status is particularly important, with invasive procedures (sentinel lymph node biopsy, axillary dissection) considered the gold standard despite being associated with surgical morbidity.
18F-FDG PET/CT is a standard tool for breast cancer staging and monitoring, but it does not automatically classify cancer subtypes. Currently, subtype identification requires invasive tumor biopsy with immunohistochemical staining, which is subject to sampling errors and cannot capture heterogeneity across multiple disease sites. There is therefore clinical interest in developing non-invasive approaches that use PET/CT imaging parameters and clinical data to predict breast cancer subtype.
This study investigated whether an artificial neural network (ANN) model trained on clinical variables and tumor marker values collected during routine 18F-FDG PET/CT workups could accurately classify patients into breast cancer subtypes, potentially enabling non-invasive subtype prediction as part of the existing imaging evaluation.
The study enrolled 142 patients who underwent 18F-FDG PET/CT at a nuclear medicine facility in Bangladesh. The median patient age was 50 years (range 26-73). Confirmed histological subtypes were: invasive ductal carcinoma (IDC, n=119), invasive lobular carcinoma (ILC, n=14), and other subtypes (n=9). The data was split 69/31% into training (98 patients) and testing (44 patients) sets using K-fold cross-validation.
PET/CT imaging followed a standard protocol: minimum 6-hour fast, blood glucose confirmed below 150 mg/dL, FDG dose of 4.07 MBq/kg injected intravenously, 1-hour uptake period before scanning from skull base to proximal thigh. CT used 100 mA and 120 kVp parameters without contrast enhancement. PET images were reconstructed using time-of-flight with point-spread-function modeling. Quantitative analysis including SUVmax measurements was performed by two nuclear medicine specialists with 5 and 12 years of experience.
The artificial neural network architecture comprised three layers: an input layer with 6 features (age, height, weight, blood glucose, injected FDG amount, and tumor marker CA-15-3 value), two hidden layers (5 and 4 neurons, sigmoid activation function), and an output layer with 3 nodes corresponding to IDC, ILC, and other subtypes using a Softmax activation function with cross-entropy loss.
The Softmax function converts the network's raw outputs into class probabilities that sum to 1, providing a probability estimate for each subtype simultaneously. For example, given an IDC case, the model might output probabilities of (0.85, 0.13, 0.02) for IDC, ILC, and other - indicating 85% confidence in the IDC prediction. Cross-entropy loss penalizes confident wrong predictions more severely than uncertain ones, guiding the model toward well-calibrated probability estimates.
The ANN model achieved an overall accuracy of 95.77% with K-fold cross-validation. The average sensitivity (true positive rate) was 0.958 and average specificity was 0.889, with an average AUC of 0.985 across all three subtype classifications. These results indicate the model can reliably distinguish between IDC, ILC, and other breast cancer subtypes from clinical data collected during PET/CT evaluation.
Subtype-specific AUC values were 0.987 for IDC, 0.964 for ILC, and 0.995 for other subtypes. The confusion matrix shows the model correctly classified all 119 IDC cases, 10 of 14 ILC cases, and 7 of 9 other cases. The 4 misclassified ILC cases (3 classified as IDC, 1 as other) and 2 misclassified other-subtype cases represent the main error sources.
Feature importance analysis revealed that the tumor marker CA-15-3 was the most significant predictor of breast cancer subtype, showing statistically significant differences (p<0.05) between subtypes. In contrast, blood glucose and injected FDG amount were non-significant differentiators between subtypes - indicating that the predictive power came primarily from the biological tumor marker rather than metabolic imaging parameters.
The performance compares favorably with related published work. An ANN model for nodal involvement prediction reported AUC 0.955 with sensitivity 0.82 and specificity 0.90; a deep learning breast cancer model achieved AUC 0.99, sensitivity 96.3%, and specificity 89.6%; and a CNN-based model for breast cancer metastasis detection achieved accuracy ranging from 86% to 89%. The current model's AUC of 0.985 is competitive with these benchmarks for a classification rather than detection task.
Artificial neural networks (ANNs) are computational models loosely inspired by biological neural networks. They consist of interconnected nodes (neurons) organized in layers, where each connection has a trainable weight that is adjusted during learning to minimize prediction error. Unlike deep convolutional neural networks that process image pixels, this ANN processes tabular clinical features, making it simpler and more interpretable.
The architecture uses sigmoid activation functions in the hidden layers, which map any input to a value between 0 and 1, introducing non-linearity that allows the network to learn complex decision boundaries. The output layer uses Softmax, which converts the three output nodes into a probability distribution summing to 1 across the three subtype classes, enabling simultaneous multi-class classification.
Training used K-fold cross-validation, a technique that divides the dataset into K subsets, trains the model K times (each time holding out one subset for testing and using the rest for training), and averages results. This approach provides more reliable performance estimates than a single train-test split, especially for smaller datasets where any particular split might not be representative.
Input features were standardized (normalized) before training to ensure that variables measured in different units (e.g., age in years vs. tumor marker in U/mL) contribute equally to the model. The Wrapper method using Shapley values through the Waikato Environment for Knowledge Analysis (WEKA) was used for feature selection to identify the most predictive input variables, ensuring the model used the most informative available data.
The finding that tumor marker CA-15-3 is the dominant predictor - rather than imaging-derived features like SUVmax - has important implications. CA-15-3 is a glycoprotein shed by breast cancer cells into the bloodstream, and its level correlates with breast cancer disease burden and type. Different subtypes produce CA-15-3 at different rates, explaining its subtype-discriminating power. This suggests that clinical blood tests, not just imaging, carry subtype-relevant information that AI can exploit.
The model's ability to distinguish ILC from IDC using clinical features is particularly notable given that ILC presents imaging challenges for FDG PET - ILC tends to have lower FDG uptake due to its diffuse growth pattern and low proliferation rate, making it harder to detect by standard PET metrics. If a clinical feature-based ANN can identify ILC cases, it could flag patients who need dedicated assessment strategies different from standard FDG PET protocols.
The proposed workflow integrates naturally into the existing clinical pathway: when a patient undergoes PET/CT for breast cancer staging, their clinical data (age, weight, blood glucose, FDG dose, CA-15-3) is already collected as part of standard pre-scan evaluation. Running this data through an ANN model would add no additional burden to the patient or clinical workflow while potentially providing subtype prediction to complement or guide biopsy interpretation.
Compared to invasive biopsy with immunohistochemistry, an ANN-based approach would have practical limitations. The model cannot replace pathological confirmation, but it could serve as a decision support tool for triaging patients, planning biopsy sites, or flagging diagnostic uncertainty. For patients with multiple lesions, where biopsy sampling errors are a concern, a whole-body perspective from PET/CT combined with ANN subtype prediction could provide complementary information about the dominant tumor biology.
This study demonstrates that an ANN model using six clinical variables collected during routine 18F-FDG PET/CT evaluation can predict breast cancer histological subtype (IDC, ILC, or other) with 95.77% accuracy and AUC 0.985. The results support the feasibility of incorporating ANN-based subtype prediction into the PET/CT clinical workflow as a non-invasive complement to tissue biopsy.
The finding that tumor marker CA-15-3 is the most important feature for subtype discrimination suggests that future models combining CA-15-3 with PET-derived radiomic features or additional biomarkers could further improve subtype prediction accuracy. Multi-modal models incorporating imaging features alongside clinical data may better capture the full spectrum of subtype-relevant biological variation.
Key limitations include the single-center design with a small sample size (142 patients), imbalanced subtype representation (119 IDC vs. 14 ILC vs. 9 other), and lack of external validation. The model was trained and tested on data from a single institution in Bangladesh, which may limit generalizability across different patient populations and healthcare settings.
Future work should validate the model on larger multicenter datasets with more balanced subtype representation, including rarer subtypes that were grouped together as "other" in this study. Integration with genomic and radiomic data, along with prospective clinical evaluation, would be required before clinical implementation. Ultimately, a validated ANN subtype prediction tool could guide personalized treatment planning decisions earlier in the diagnostic pathway, potentially before biopsy results are available.